Merge branch 'master' into 20260322a_ltx_contextwin

.github/workflows/detect-unreviewed-merge.yml

@@ -0,0 +1,24 @@
+name: Detect Unreviewed Merge
+
+# SOC 2 compliance — reusable workflow lives in Comfy-Org/github-workflows,
+# tracking issues are filed in Comfy-Org/unreviewed-merges.
+
+on:
+  push:
+    branches: [master]
+
+concurrency:
+  group: detect-unreviewed-merge-${{ github.sha }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+  pull-requests: read
+
+jobs:
+  detect:
+    uses: Comfy-Org/github-workflows/.github/workflows/detect-unreviewed-merge.yml@4d9cb6b87f953bb7cd69954280e1465fb9bd2040 # v1
+    with:
+      approval-mode: latest-per-reviewer
+    secrets:
+      UNREVIEWED_MERGES_TOKEN: ${{ secrets.UNREVIEWED_MERGES_TOKEN }}

comfy/bg_removal_model.py

@@ -55,12 +55,7 @@ class BackgroundRemovalModel():
         out = torch.nn.functional.interpolate(out, size=(H, W), mode="bicubic", antialias=False)
 
         mask = out.sigmoid().to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
-        if mask.ndim == 3:
-            mask = mask.unsqueeze(0)
-        if mask.shape[1] != 1:
-            mask = mask.movedim(-1, 1)
-
-        return mask
+        return mask.squeeze(1)  # (B, 1, H, W) -> (B, H, W)
 
 
 def load_background_removal_model(sd):

comfy/cli_args.py

@@ -149,6 +149,7 @@ parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=Non
 parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.")
 parser.add_argument("--disable-dynamic-vram", action="store_true", help="Disable dynamic VRAM and use estimate based model loading.")
 parser.add_argument("--enable-dynamic-vram", action="store_true", help="Enable dynamic VRAM on systems where it's not enabled by default.")
+parser.add_argument("--fast-disk", action="store_true", help="Prefer disk-backed dynamic loading and offload over unpinned RAM. Can be faster for users with fast NVME disks.")
 
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 

comfy/float.py

@@ -1,5 +1,20 @@
+import logging
+
 import torch
 
+_CK_STOCHASTIC_ROUNDING_AVAILABLE = False
+try:
+    import comfy_kitchen as ck
+    _ck_stochastic_rounding_fp8 = ck.stochastic_rounding_fp8
+    _CK_STOCHASTIC_ROUNDING_AVAILABLE = True
+except (AttributeError, ImportError):
+    logging.warning("comfy_kitchen does not support stochastic FP8 rounding, please update comfy_kitchen.")
+
+if not _CK_STOCHASTIC_ROUNDING_AVAILABLE:
+    def _ck_stochastic_rounding_fp8(value, rng, dtype):
+        raise NotImplementedError("comfy_kitchen does not support stochastic FP8 rounding")
+
+
 def calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=None):
     mantissa_scaled = torch.where(
         normal_mask,
@@ -57,6 +72,10 @@ def stochastic_rounding(value, dtype, seed=0):
     if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2:
         generator = torch.Generator(device=value.device)
         generator.manual_seed(seed)
+        if _CK_STOCHASTIC_ROUNDING_AVAILABLE:
+            rng = torch.randint(0, 256, value.size(), dtype=torch.uint8, layout=value.layout, device=value.device, generator=generator)
+            return _ck_stochastic_rounding_fp8(value, rng, dtype)
+
         output = torch.empty_like(value, dtype=dtype)
         num_slices = max(1, (value.numel() / (4096 * 4096)))
         slice_size = max(1, round(value.shape[0] / num_slices))

comfy/latent_formats.py

@@ -799,13 +799,15 @@ class ZImagePixelSpace(ChromaRadiance):
     """
     pass
 
-
 class HiDreamO1Pixel(ChromaRadiance):
     """Pixel-space latent format for HiDream-O1.
     No VAE — model patches/unpatches raw RGB internally with patch_size=32.
     """
     pass
 
+class PixelDiTPixel(ChromaRadiance):
+    pass
+
 class CogVideoX(LatentFormat):
     """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b).
 

comfy/ldm/audio/dit.py

@@ -433,11 +433,11 @@ class Attention(nn.Module):
         if self.differential:
             q, q_diff = q.unbind(dim=1)
             k, k_diff = k.unbind(dim=1)
-            out      = optimized_attention(q,      k,      v, h, skip_reshape=True, transformer_options=transformer_options)
-            out_diff = optimized_attention(q_diff, k_diff, v, h, skip_reshape=True, transformer_options=transformer_options)
+            out      = optimized_attention(q,      k,      v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options)
+            out_diff = optimized_attention(q_diff, k_diff, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options)
             out = out - out_diff
         else:
-            out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options)
+            out = optimized_attention(q, k, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options)
 
         out = self.to_out(out)
 

comfy/ldm/audio/vae_sa3.py

@@ -138,11 +138,11 @@ class Attention(nn.Module):
                 k_diff = _apply_rotary_pos_emb(k_diff.float(), freqs).to(k_dtype)
 
         if self.differential:
-            out = (optimized_attention(q, k, v, h, mask=mask, skip_reshape=True)
-                   - optimized_attention(q_diff, k_diff, v, h, mask=mask, skip_reshape=True))
+            out = (optimized_attention(q, k, v, h, mask=mask, skip_reshape=True, low_precision_attention=False)
+                   - optimized_attention(q_diff, k_diff, v, h, mask=mask, skip_reshape=True, low_precision_attention=False))
             del q, k, v, q_diff, k_diff
         else:
-            out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True)
+            out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True, low_precision_attention=False)
             del q, k, v
 
         return self.to_out(out)

comfy/ldm/cosmos/predict2.py

@@ -14,15 +14,7 @@ from torchvision import transforms
 import comfy.patcher_extension
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.ldm.common_dit
-
-def apply_rotary_pos_emb(
-    t: torch.Tensor,
-    freqs: torch.Tensor,
-) -> torch.Tensor:
-    t_ = t.reshape(*t.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).float()
-    t_out = freqs[..., 0] * t_[..., 0] + freqs[..., 1] * t_[..., 1]
-    t_out = t_out.movedim(-1, -2).reshape(*t.shape).type_as(t)
-    return t_out
+import comfy.quant_ops
 
 
 # ---------------------- Feed Forward Network -----------------------
@@ -173,8 +165,7 @@ class Attention(nn.Module):
             k = self.k_norm(k)
             v = self.v_norm(v)
             if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
-                q = apply_rotary_pos_emb(q, rope_emb)
-                k = apply_rotary_pos_emb(k, rope_emb)
+                q, k = comfy.quant_ops.ck.apply_rope_split_half(q, k, rope_emb)
             return q, k, v
 
         q, k, v = apply_norm_and_rotary_pos_emb(q, k, v, rope_emb)

comfy/ldm/ernie/model.py

@@ -5,6 +5,7 @@ import torch.nn.functional as F
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
+import comfy.quant_ops
 
 def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
     assert dim % 2 == 0
@@ -19,15 +20,6 @@ def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
     out = torch.stack([torch.cos(out), torch.sin(out)], dim=0)
     return out.to(dtype=torch.float32, device=pos.device)
 
-def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
-    rot_dim = freqs_cis.shape[-1]
-    x, x_pass = x_in[..., :rot_dim], x_in[..., rot_dim:]
-    cos_ = freqs_cis[0]
-    sin_ = freqs_cis[1]
-    x1, x2 = x.chunk(2, dim=-1)
-    x_rotated = torch.cat((-x2, x1), dim=-1)
-    return torch.cat((x * cos_ + x_rotated * sin_, x_pass), dim=-1)
-
 class ErnieImageEmbedND3(nn.Module):
     def __init__(self, dim: int, theta: int, axes_dim: tuple):
         super().__init__()
@@ -37,8 +29,16 @@ class ErnieImageEmbedND3(nn.Module):
 
     def forward(self, ids: torch.Tensor) -> torch.Tensor:
         emb = torch.cat([rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(3)], dim=-1)
-        emb = emb.unsqueeze(3)  # [2, B, S, 1, head_dim//2]
-        return torch.stack([emb, emb], dim=-1).reshape(*emb.shape[:-1], -1)  # [B, S, 1, head_dim]
+        cos_ = emb[0]
+        sin_ = emb[1]
+        N = cos_.shape[-1]
+        half = N // 2
+        cos_top = cos_[..., :half].repeat_interleave(2, dim=-1)
+        sin_top = sin_[..., :half].repeat_interleave(2, dim=-1)
+        cos_bot = cos_[..., half:].repeat_interleave(2, dim=-1)
+        sin_bot = sin_[..., half:].repeat_interleave(2, dim=-1)
+        rot = torch.stack([cos_top, -sin_top, sin_bot, cos_bot], dim=-1)
+        return rot.reshape(*rot.shape[:-1], 2, 2).unsqueeze(2)
 
 class ErnieImagePatchEmbedDynamic(nn.Module):
     def __init__(self, in_channels: int, embed_dim: int, patch_size: int, operations, device=None, dtype=None):
@@ -115,8 +115,7 @@ class ErnieImageAttention(nn.Module):
         key = self.norm_k(key)
 
         if image_rotary_emb is not None:
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
+            query, key = comfy.quant_ops.ck.apply_rope_split_half(query, key, image_rotary_emb)
 
         q_flat = query.reshape(B, S, -1)
         k_flat = key.reshape(B, S, -1)
@@ -274,7 +273,7 @@ class ErnieImageModel(nn.Module):
 
         image_ids = image_ids.view(1, N_img, 3).expand(B, -1, -1)
 
-        rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1)).to(x.dtype)
+        rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1))
         del image_ids, text_ids
 
         sample = self.time_proj(timesteps).to(dtype)

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -211,7 +211,7 @@ class TimestepEmbedder(nn.Module):
     Embeds scalar timesteps into vector representations.
     """
 
-    def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None, max_period=10000):
         super().__init__()
         if output_size is None:
             output_size = hidden_size
@@ -221,9 +221,10 @@ class TimestepEmbedder(nn.Module):
             operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device),
         )
         self.frequency_embedding_size = frequency_embedding_size
+        self.max_period = max_period
 
     def forward(self, t, dtype, **kwargs):
-        t_freq = timestep_embedding(t, self.frequency_embedding_size).to(dtype)
+        t_freq = timestep_embedding(t, self.frequency_embedding_size, max_period=self.max_period).to(dtype)
         t_emb = self.mlp(t_freq)
         return t_emb
 

comfy/ldm/pixeldit/model.py

@@ -0,0 +1,239 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ldm.common_dit
+import comfy.patcher_extension
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.hidream.model import FeedForwardSwiGLU
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
+
+from .modules import (
+    FinalLayer,
+    PatchTokenEmbedder,
+    PiTBlock,
+    PixelTokenEmbedder,
+    apply_adaln_,
+    precompute_freqs_cis_2d,
+)
+
+
+class MMDiTJointAttention(nn.Module):
+    """Joint MMDiT attention with separate Q/K/V/proj for image and text streams.
+
+    RoPE is applied to each stream before concatenation so each stream uses its own
+    2D/1D positional encoding. Concat order is [text, image] (text first).
+    """
+    def __init__(self, dim, num_heads=8, qkv_bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.qkv_x = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.qkv_y = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+
+        self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+
+        self.proj_x = operations.Linear(dim, dim, dtype=dtype, device=device)
+        self.proj_y = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+    def forward(self, x, y, pos_img, pos_txt=None, attn_mask=None, transformer_options={}):
+        B, Nx, _ = x.shape
+        _, Ny, _ = y.shape
+        H = self.num_heads
+        D = self.head_dim
+
+        qkv_x = self.qkv_x(x).reshape(B, Nx, 3, H, D).permute(2, 0, 3, 1, 4)
+        qx, kx, vx = qkv_x.unbind(0)
+        qx = self.q_norm_x(qx)
+        kx = self.k_norm_x(kx)
+
+        qkv_y = self.qkv_y(y).reshape(B, Ny, 3, H, D).permute(2, 0, 3, 1, 4)
+        qy, ky, vy = qkv_y.unbind(0)
+        qy = self.q_norm_y(qy)
+        ky = self.k_norm_y(ky)
+
+        qx, kx = apply_rope(qx, kx, pos_img[None, None])
+        if pos_txt is not None:
+            qy, ky = apply_rope(qy, ky, pos_txt[None, None])
+
+        q_joint = torch.cat([qy, qx], dim=2)
+        k_joint = torch.cat([ky, kx], dim=2)
+        v_joint = torch.cat([vy, vx], dim=2)
+
+        out_joint = optimized_attention(
+            q_joint, k_joint, v_joint, H,
+            mask=attn_mask, skip_reshape=True, skip_output_reshape=True,
+            transformer_options=transformer_options,
+        )
+
+        out_y = out_joint[:, :, :Ny, :].transpose(1, 2).reshape(B, Ny, H * D)
+        out_x = out_joint[:, :, Ny:, :].transpose(1, 2).reshape(B, Nx, H * D)
+
+        return self.proj_x(out_x), self.proj_y(out_y)
+
+
+class MMDiTBlockT2I(nn.Module):
+    def __init__(self, hidden_size, groups, mlp_ratio=4.0, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_x1 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.norm_y1 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.attn = MMDiTJointAttention(hidden_size, num_heads=groups, qkv_bias=False, dtype=dtype, device=device, operations=operations)
+        self.norm_x2 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.norm_y2 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.mlp_x = FeedForwardSwiGLU(hidden_size, mlp_hidden_dim, multiple_of=1, dtype=dtype, device=device, operations=operations)
+        self.mlp_y = FeedForwardSwiGLU(hidden_size, mlp_hidden_dim, multiple_of=1, dtype=dtype, device=device, operations=operations)
+        self.adaLN_modulation_img = nn.Sequential(operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device))
+        self.adaLN_modulation_txt = nn.Sequential(operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device))
+
+    def forward(self, x, y, c, pos_img, pos_txt=None, attn_mask=None, transformer_options={}):
+        shift_msa_x, scale_msa_x, gate_msa_x, shift_mlp_x, scale_mlp_x, gate_mlp_x = self.adaLN_modulation_img(c).chunk(6, dim=-1)
+        shift_msa_y, scale_msa_y, gate_msa_y, shift_mlp_y, scale_mlp_y, gate_mlp_y = self.adaLN_modulation_txt(c).chunk(6, dim=-1)
+
+        x_norm = apply_adaln_(self.norm_x1(x), shift_msa_x, scale_msa_x)
+        y_norm = apply_adaln_(self.norm_y1(y), shift_msa_y, scale_msa_y)
+        attn_x, attn_y = self.attn(x_norm, y_norm, pos_img, pos_txt, attn_mask, transformer_options=transformer_options)
+        x = torch.addcmul(x, gate_msa_x, attn_x)
+        y = torch.addcmul(y, gate_msa_y, attn_y)
+
+        x = torch.addcmul(x, gate_mlp_x, self.mlp_x(apply_adaln_(self.norm_x2(x), shift_mlp_x, scale_mlp_x)))
+        y = torch.addcmul(y, gate_mlp_y, self.mlp_y(apply_adaln_(self.norm_y2(y), shift_mlp_y, scale_mlp_y)))
+        return x, y
+
+
+class PixDiT_T2I(nn.Module):
+    """PixelDiT T2I model. Hardcoded for the released 1024px Stage-3 checkpoint
+    (also runs at 512px when fed the appropriate latent size and flow_shift).
+
+    Forward:
+      x:        [B, 3, H, W] pixel-space input (no VAE)
+      timesteps:[B] in [0, 1000] (ComfyUI flow sampling convention)
+      context:  [B, Ltxt, 2304] Gemma-2-2b-it hidden states (chi_prompt prepended)
+    Returns flow-matching velocity [B, 3, H, W].
+    """
+    def __init__(
+        self,
+        in_channels=3,
+        num_groups=24,
+        hidden_size=1536,
+        pixel_hidden_size=16,
+        pixel_attn_hidden_size=1152,
+        pixel_num_groups=16,
+        patch_depth=14,
+        pixel_depth=2,
+        patch_size=16,
+        txt_embed_dim=2304,
+        txt_max_length=300,
+        use_text_rope=True,
+        text_rope_theta=10000.0,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None,
+        pixel_mlp_chunks=2,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.hidden_size = hidden_size
+        self.num_groups = num_groups
+        self.patch_depth = patch_depth
+        self.pixel_depth = pixel_depth
+        self.patch_size = patch_size
+        self.pixel_hidden_size = pixel_hidden_size
+        self.pixel_attn_hidden_size = pixel_attn_hidden_size
+        self.pixel_num_groups = pixel_num_groups
+        self.txt_embed_dim = txt_embed_dim
+        self.txt_max_length = txt_max_length
+        self.use_text_rope = use_text_rope
+        self.text_rope_theta = text_rope_theta
+
+        self.pixel_embedder = PixelTokenEmbedder(self.in_channels, self.pixel_hidden_size, dtype=dtype, device=device, operations=operations)
+        self.s_embedder = PatchTokenEmbedder(self.in_channels * self.patch_size ** 2, self.hidden_size, bias=True, dtype=dtype, device=device, operations=operations)
+        self.t_embedder = TimestepEmbedder(self.hidden_size, dtype=dtype, device=device, operations=operations, max_period=10)
+        self.y_embedder = PatchTokenEmbedder(self.txt_embed_dim, self.hidden_size, bias=True, use_norm=True, dtype=dtype, device=device, operations=operations)
+        self.y_pos_embedding = nn.Parameter(torch.empty(1, self.txt_max_length, self.hidden_size, dtype=dtype, device=device))
+
+        self.patch_blocks = nn.ModuleList([
+            MMDiTBlockT2I(self.hidden_size, self.num_groups,
+                          dtype=dtype, device=device, operations=operations)
+            for _ in range(self.patch_depth)
+        ])
+        self.pixel_blocks = nn.ModuleList([
+            PiTBlock(
+                self.pixel_hidden_size,
+                self.hidden_size,
+                patch_size=self.patch_size,
+                num_heads=self.num_groups,
+                attn_hidden_size=self.pixel_attn_hidden_size,
+                attn_num_heads=self.pixel_num_groups,
+                dtype=dtype, device=device, operations=operations,
+                mlp_chunks=pixel_mlp_chunks,
+            )
+            for _ in range(self.pixel_depth)
+        ])
+
+        self.final_layer = FinalLayer(self.pixel_hidden_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+    def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts):
+        return precompute_freqs_cis_2d(self.hidden_size // self.num_groups, height, width, device=device, dtype=dtype, **rope_opts)
+
+    def _fetch_text_pos(self, length, device, dtype):
+        return rope(torch.arange(length, dtype=torch.float32, device=device).reshape(1, -1), self.hidden_size // self.num_groups, self.text_rope_theta).squeeze(0).to(dtype=dtype)
+
+    def forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward, self, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options),
+        ).execute(x, timesteps, context, attention_mask, transformer_options, **kwargs)
+
+    def _pre_patch_block(self, s, i, **kwargs):
+        """Hook for subclasses to inject per-block state into the patch stream (e.g. PiD's LQ gate)."""
+        return s
+
+    def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs):
+        H_orig, W_orig = x.shape[2], x.shape[3]
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        B, _, H, W = x.shape
+        Hs = H // self.patch_size
+        Ws = W // self.patch_size
+        L = Hs * Ws
+
+        pos_img = self._fetch_patch_pos(Hs, Ws, x.device, x.dtype, **(transformer_options.get("rope_options") or {}))
+        x_patches = F.unfold(x, kernel_size=self.patch_size, stride=self.patch_size).transpose(1, 2)
+
+        t_emb = self.t_embedder(timesteps.view(-1), x.dtype).view(B, -1, self.hidden_size)
+
+        if context is None or context.dim() != 3:
+            raise ValueError("PixDiT_T2I requires context (text embeddings) of shape [B, L, D]")
+        Ltxt = min(context.shape[1], self.txt_max_length)
+        y = context[:, :Ltxt, :]
+        y_emb = self.y_embedder(y).view(B, Ltxt, self.hidden_size)
+        y_emb = y_emb + self.y_pos_embedding[:, :Ltxt, :].to(y_emb) # y_pos_embedding is a raw nn.Parameter
+
+        condition = F.silu(t_emb)
+        pos_txt = self._fetch_text_pos(Ltxt, x.device, x.dtype) if self.use_text_rope else None
+
+        s = self.s_embedder(x_patches)
+        for i, blk in enumerate(self.patch_blocks):
+            s = self._pre_patch_block(s, i, **kwargs)
+            s, y_emb = blk(s, y_emb, condition, pos_img, pos_txt, None, transformer_options=transformer_options)
+        s = F.silu(t_emb + s)
+
+        s_cond = s.view(B * L, self.hidden_size)
+        x_pixels = self.pixel_embedder(x, patch_size=self.patch_size)
+        for blk in self.pixel_blocks:
+            x_pixels = blk(x_pixels, s_cond, H, W, self.patch_size, mask=None, transformer_options=transformer_options)
+
+        x_pixels = self.final_layer(x_pixels)
+        C_out = self.out_channels
+        P2 = self.patch_size * self.patch_size
+        x_pixels = x_pixels.view(B, L, P2, C_out).permute(0, 3, 2, 1).reshape(B, C_out * P2, L)
+        out = F.fold(x_pixels, (H, W), kernel_size=self.patch_size, stride=self.patch_size)
+        return out[:, :, :H_orig, :W_orig]

comfy/ldm/pixeldit/modules.py

@@ -0,0 +1,187 @@
+import torch
+import torch.nn as nn
+
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, get_1d_sincos_pos_embed_from_grid_torch
+
+
+def apply_adaln_(x, shift, scale):
+    return x.addcmul_(x, scale).add_(shift)
+
+
+def precompute_freqs_cis_2d(dim, height, width, theta=10000.0, scale=16.0,
+                            ref_grid_h=None, ref_grid_w=None,
+                            scale_x=1.0, scale_y=1.0, shift_x=0.0, shift_y=0.0,
+                            device=None, dtype=torch.float32, **kwargs):
+    """2D RoPE with x/y axis frequencies interleaved at stride 2 across head dim.
+
+    rope_options:
+      scale_x / scale_y multiply the position range (RoPE extrapolation).
+      shift_x / shift_y offset the position origin (tiled / regional inference).
+    With ref_grid_h/w set, also applies NTK-aware per-axis theta scaling
+    (rope_mode='ntk_aware'): theta_axis = theta * (current/ref)^(dim_axis/(dim_axis-2)).
+    Returns Flux-format rotation matrices of shape [H*W, dim/2, 2, 2].
+    Layout of head-dim pairs: [x_0, y_0, x_1, y_1, ..., x_{dim/4-1}, y_{dim/4-1}].
+    """
+    dim_axis = dim // 2
+    if ref_grid_h is not None and dim_axis > 2:
+        h_ntk = (height / ref_grid_h) ** (dim_axis / (dim_axis - 2))
+        w_ntk = (width / ref_grid_w) ** (dim_axis / (dim_axis - 2))
+    else:
+        h_ntk = w_ntk = 1.0
+
+    x_lin = torch.linspace(shift_x, scale * scale_x + shift_x, width, device=device)
+    y_lin = torch.linspace(shift_y, scale * scale_y + shift_y, height, device=device)
+    y_grid, x_grid = torch.meshgrid(y_lin, x_lin, indexing="ij")
+    x_rope = rope(x_grid.reshape(1, -1), dim_axis, theta * w_ntk).squeeze(0)
+    y_rope = rope(y_grid.reshape(1, -1), dim_axis, theta * h_ntk).squeeze(0)
+    out = torch.stack([x_rope, y_rope], dim=2).reshape(height * width, dim // 2, 2, 2)
+    return out.to(dtype=dtype)
+
+
+def get_2d_sincos_pos_embed(embed_dim, height, width, device=None, dtype=torch.float32):
+    """Standard 2D sin/cos absolute positional embedding (ViT-style).
+
+    first half encodes W-coordinates, second half H.
+    """
+    assert embed_dim % 4 == 0
+    grid_h = torch.arange(height, dtype=torch.float32, device=device)
+    grid_w = torch.arange(width, dtype=torch.float32, device=device)
+    grid_y, grid_x = torch.meshgrid(grid_h, grid_w, indexing="ij")
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_x.reshape(-1), device=device)
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_y.reshape(-1), device=device)
+    return torch.cat([emb_w, emb_h], dim=1).to(dtype=dtype)
+
+
+class RotaryAttention(nn.Module):
+    """Single-stream self-attention with rotary positional encoding (used inside PiTBlock)."""
+    def __init__(self, dim, num_heads=8, qkv_bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+    def forward(self, x, pos, mask=None, transformer_options={}):
+        B, N, C = x.shape
+        H = self.num_heads
+        D = self.head_dim
+        qkv = self.qkv(x).reshape(B, N, 3, H, D).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)
+        q, k = apply_rope(self.q_norm(q), self.k_norm(k), pos[None, None])
+        x = optimized_attention(q, k, v, H, mask=mask, skip_reshape=True, transformer_options=transformer_options)
+        return self.proj(x)
+
+
+class FinalLayer(nn.Module):
+    def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.linear(self.norm(x))
+
+
+class PatchTokenEmbedder(nn.Module):
+    """Linear projection used both for patchified-image tokens and text-feature tokens."""
+    def __init__(self, in_chans, embed_dim, use_norm=False, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.proj = operations.Linear(in_chans, embed_dim, bias=bias, dtype=dtype, device=device)
+        self.norm = operations.RMSNorm(embed_dim, eps=1e-6, dtype=dtype, device=device) if use_norm else nn.Identity()
+
+    def forward(self, x):
+        return self.norm(self.proj(x))
+
+
+class PixelTokenEmbedder(nn.Module):
+    """Pixel-level embedder: lifts each RGB pixel to hidden_size and packs into per-patch sequences."""
+    def __init__(self, in_channels, hidden_size_output, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_size_output = hidden_size_output
+        self.proj = operations.Linear(self.in_channels, self.hidden_size_output, bias=True, dtype=dtype, device=device)
+
+    def forward(self, inputs, patch_size):
+        B, _, H, W = inputs.shape
+        Hs, Ws = H // patch_size, W // patch_size
+        P2 = patch_size * patch_size
+        x = inputs.permute(0, 2, 3, 1).contiguous()
+        x = self.proj(x)
+        pos_full = get_2d_sincos_pos_embed(self.hidden_size_output, H, W, device=x.device, dtype=x.dtype).view(H, W, self.hidden_size_output)
+        x = x + pos_full.unsqueeze(0)
+        x = x.view(B, Hs, patch_size, Ws, patch_size, self.hidden_size_output)
+        return x.permute(0, 1, 3, 2, 4, 5).reshape(B * Hs * Ws, P2, self.hidden_size_output)
+
+
+class PiTBlock(nn.Module):
+    """Pixel-level transformer block.
+
+    Compresses each patch's P^2 pixel tokens → 1 attention token via a linear,
+    runs global self-attention across patches with 2D RoPE, then expands back to P^2 tokens.
+    Conditioning is per-pixel adaLN from the patch-level features.
+    """
+    def __init__(self, pixel_hidden_size, patch_hidden_size, patch_size, num_heads, mlp_ratio=4.0,
+                 attn_hidden_size=None, attn_num_heads=None, dtype=None, device=None, operations=None, mlp_chunks=1):
+        super().__init__()
+        self.pixel_dim = pixel_hidden_size
+        self.context_dim = patch_hidden_size
+        self.attn_dim = attn_hidden_size if attn_hidden_size is not None else patch_hidden_size
+        self.num_heads = attn_num_heads if attn_num_heads is not None else num_heads
+        assert self.attn_dim % self.num_heads == 0
+
+        p2 = patch_size * patch_size
+        self.compress_to_attn = operations.Linear(p2 * self.pixel_dim, self.attn_dim, bias=True, dtype=dtype, device=device)
+        self.expand_from_attn = operations.Linear(self.attn_dim, p2 * self.pixel_dim, bias=True, dtype=dtype, device=device)
+
+        self.norm1 = operations.RMSNorm(self.pixel_dim, eps=1e-6, dtype=dtype, device=device)
+        self.attn = RotaryAttention(self.attn_dim, num_heads=self.num_heads, qkv_bias=False, dtype=dtype, device=device, operations=operations)
+        self.norm2 = operations.RMSNorm(self.pixel_dim, eps=1e-6, dtype=dtype, device=device)
+        self.mlp = Mlp(self.pixel_dim, hidden_features=int(self.pixel_dim * mlp_ratio), dtype=dtype, device=device, operations=operations)
+
+        self.adaLN_modulation_msa = operations.Linear(self.context_dim, 3 * self.pixel_dim * p2, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation_mlp = operations.Linear(self.context_dim, 3 * self.pixel_dim * p2, bias=True, dtype=dtype, device=device)
+
+        self._rope_fn = precompute_freqs_cis_2d
+        self.mlp_chunks = max(1, int(mlp_chunks))
+
+    def _fetch_pos(self, height, width, device, dtype, **rope_opts):
+        return self._rope_fn(self.attn_dim // self.num_heads, height, width, device=device, dtype=dtype, **rope_opts)
+
+    def forward(self, x, s_cond, image_height, image_width, patch_size, mask=None, transformer_options={}):
+        BL, P2, _ = x.shape
+        Hs, Ws = image_height // patch_size, image_width // patch_size
+        L = Hs * Ws
+        B = BL // L
+
+        # Attention path uses only msa params; compute, use, free before mlp params allocate.
+        msa_params = self.adaLN_modulation_msa(s_cond).view(BL, P2, 3 * self.pixel_dim)
+        shift_msa, scale_msa, gate_msa = msa_params.chunk(3, dim=-1)
+
+        x_norm = apply_adaln_(self.norm1(x), shift_msa, scale_msa)
+        x_flat = x_norm.view(BL, P2 * self.pixel_dim)
+
+        x_comp = self.compress_to_attn(x_flat).view(B, L, self.attn_dim)
+        pos_comp = self._fetch_pos(Hs, Ws, x.device, x.dtype, **(transformer_options.get("rope_options") or {}))
+        attn_out = self.attn(x_comp, pos_comp, mask=mask, transformer_options=transformer_options)
+        attn_flat = self.expand_from_attn(attn_out.view(B * L, self.attn_dim))
+        attn_exp = attn_flat.view(BL, P2, self.pixel_dim)
+        x = torch.addcmul(x, gate_msa, attn_exp)
+        del msa_params, shift_msa, scale_msa, gate_msa
+
+        mlp_params = self.adaLN_modulation_mlp(s_cond).view(BL, P2, 3 * self.pixel_dim)
+        shift_mlp, scale_mlp, gate_mlp = mlp_params.chunk(3, dim=-1)
+        gate_mlp = gate_mlp.contiguous()  # detach from mlp_params so the del below frees shift+scale storage before the MLP
+        mlp_input = apply_adaln_(self.norm2(x), shift_mlp, scale_mlp)
+        del mlp_params, shift_mlp, scale_mlp
+
+        # MLP in chunks since the peak memory usage is huge here
+        chunk_size = (BL + self.mlp_chunks - 1) // self.mlp_chunks
+        for s in range(0, BL, chunk_size):
+            e = min(s + chunk_size, BL)
+            x[s:e].addcmul_(gate_mlp[s:e], self.mlp(mlp_input[s:e]))
+        return x

comfy/ldm/pixeldit/pid.py

@@ -0,0 +1,227 @@
+"""PiD — Pixel Diffusion Decoder. Decodes a Flux/SD3/Flux2/Z-Image latent
+directly to a 4x-upscaled image in 4 distilled flow-matching steps. PixDiT_T2I
+body + LQ projection branch injected before each MMDiT patch block.
+"""
+
+from typing import List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .model import PixDiT_T2I
+from .modules import precompute_freqs_cis_2d
+
+
+class SigmaAwareGatePerTokenPerDim(nn.Module):
+    """gate = sigmoid(content_proj(cat[x, lq]) - exp(log_alpha) * sigma); out = x + gate * lq.
+
+    Trained init gives ~0.88 gate at sigma=0, ~0.05 at sigma=1.
+    """
+
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.content_proj = operations.Linear(dim * 2, dim, dtype=dtype, device=device)
+        self.log_alpha = nn.Parameter(torch.empty((), dtype=dtype, device=device))
+
+    def forward(self, x: torch.Tensor, lq: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor:
+        content_logit = self.content_proj(torch.cat([x, lq], dim=-1))
+        # log_alpha is a raw nn.Parameter -> doesn't auto-cast under dynamic VRAM.
+        log_alpha = self.log_alpha.to(device=x.device, dtype=torch.float32)
+        sigma_offset = -log_alpha.exp() * sigma.float().view(-1, 1, 1)
+        gate = torch.sigmoid(content_logit + sigma_offset)
+        return x + (gate * lq).to(x.dtype)
+
+
+class ResBlock(nn.Module):
+    """Pre-activation ResNet block: GN -> SiLU -> Conv -> GN -> SiLU -> Conv + skip."""
+
+    def __init__(self, channels: int, num_groups: int = 4, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.block = nn.Sequential(
+            operations.GroupNorm(num_groups, channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(channels, channels, kernel_size=3, padding=1, dtype=dtype, device=device),
+            operations.GroupNorm(num_groups, channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(channels, channels, kernel_size=3, padding=1, dtype=dtype, device=device),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+
+class LQProjection2D(nn.Module):
+    """LQ latent -> per-block patch-aligned features for controlnet-style injection."""
+
+    def __init__(
+        self,
+        latent_channels: int,
+        hidden_dim: int = 512,
+        out_dim: int = 1536,
+        patch_size: int = 16,
+        sr_scale: int = 4,
+        latent_spatial_down_factor: int = 8,
+        num_res_blocks: int = 4,
+        num_outputs: int = 7,
+        interval: int = 2,
+        dtype=None, device=None, operations=None,
+    ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.hidden_dim = hidden_dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.sr_scale = sr_scale
+        self.latent_spatial_down_factor = latent_spatial_down_factor
+        self.num_outputs = num_outputs
+        self.interval = interval
+
+        z_to_patch_ratio = (sr_scale * latent_spatial_down_factor) / patch_size
+        self.z_to_patch_ratio = z_to_patch_ratio
+        if z_to_patch_ratio >= 1:
+            self.latent_fold_factor = 0
+            latent_proj_in_ch = latent_channels
+        else:
+            fold_factor = int(1 / z_to_patch_ratio)
+            assert fold_factor * z_to_patch_ratio == 1.0
+            self.latent_fold_factor = fold_factor
+            latent_proj_in_ch = latent_channels * fold_factor * fold_factor
+
+        layers = [
+            operations.Conv2d(latent_proj_in_ch, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(hidden_dim, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device),
+        ]
+        for _ in range(num_res_blocks):
+            layers.append(ResBlock(hidden_dim, dtype=dtype, device=device, operations=operations))
+        self.latent_proj = nn.Sequential(*layers)
+
+        self.output_heads = nn.ModuleList(
+            [operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) for _ in range(num_outputs)]
+        )
+        self.gate_modules = nn.ModuleList(
+            [SigmaAwareGatePerTokenPerDim(out_dim, dtype=dtype, device=device, operations=operations)
+             for _ in range(num_outputs)]
+        )
+
+    def is_gate_active(self, block_idx: int) -> bool:
+        return block_idx % self.interval == 0
+
+    def output_index(self, block_idx: int) -> int:
+        return block_idx // self.interval
+
+    def gate(self, x: torch.Tensor, lq_feature: torch.Tensor, sigma: torch.Tensor, out_idx: int) -> torch.Tensor:
+        return self.gate_modules[out_idx](x, lq_feature, sigma)
+
+    def _align_latent_to_patch_grid(self, lq_latent: torch.Tensor, pH: int, pW: int) -> torch.Tensor:
+        B, z_dim = lq_latent.shape[:2]
+        if self.z_to_patch_ratio >= 1:
+            if lq_latent.shape[2] != pH or lq_latent.shape[3] != pW:
+                z_aligned = F.interpolate(lq_latent, size=(pH, pW), mode="nearest")
+            else:
+                z_aligned = lq_latent
+        else:
+            f = self.latent_fold_factor
+            zH_expected, zW_expected = pH * f, pW * f
+            if lq_latent.shape[2] != zH_expected or lq_latent.shape[3] != zW_expected:
+                lq_latent = F.interpolate(lq_latent, size=(zH_expected, zW_expected), mode="nearest")
+            z_aligned = lq_latent.reshape(B, z_dim, pH, f, pW, f).permute(0, 1, 3, 5, 2, 4)
+            z_aligned = z_aligned.reshape(B, z_dim * f * f, pH, pW)
+        return self.latent_proj(z_aligned)
+
+    def forward(self, lq_latent: torch.Tensor, target_pH: int, target_pW: int) -> List[torch.Tensor]:
+        feat = self._align_latent_to_patch_grid(lq_latent, target_pH, target_pW)
+        B, C, H, W = feat.shape
+        tokens = feat.permute(0, 2, 3, 1).contiguous().view(B, H * W, C)
+        return [head(tokens) for head in self.output_heads]
+
+
+class PidNet(PixDiT_T2I):
+    """PixDiT_T2I + LQ injection (one sigma-gated feature inserted before each patch block)."""
+
+    def __init__(
+        self,
+        lq_latent_channels: int = 16,
+        lq_hidden_dim: int = 512,
+        lq_num_res_blocks: int = 4,
+        lq_interval: int = 2,
+        sr_scale: int = 4,
+        latent_spatial_down_factor: int = 8,
+        rope_ref_h: int = 1024, # NTK ref resolution in PIXEL units: 1024px / patch=16 -> grid_ref=64.
+        rope_ref_w: int = 1024,
+        image_model=None,
+        dtype=None, device=None, operations=None,
+        **pixdit_kwargs,
+    ):
+        super().__init__(dtype=dtype, device=device, operations=operations, **pixdit_kwargs)
+
+        self.rope_ref_grid_h = rope_ref_h // self.patch_size
+        self.rope_ref_grid_w = rope_ref_w // self.patch_size
+
+        # Parent's PiTBlocks were built with plain RoPE — swap in NTK-aware.
+        def _pit_rope_fn(head_dim, h, w, device=None, dtype=torch.float32, **rope_opts):
+            return precompute_freqs_cis_2d(head_dim, h, w, ref_grid_h=self.rope_ref_grid_h, ref_grid_w=self.rope_ref_grid_w, device=device, dtype=dtype, **rope_opts)
+        for blk in self.pixel_blocks:
+            blk._rope_fn = _pit_rope_fn
+
+        num_lq_outputs = (self.patch_depth + lq_interval - 1) // lq_interval
+        self.lq_proj = LQProjection2D(
+            latent_channels=lq_latent_channels,
+            hidden_dim=lq_hidden_dim,
+            out_dim=self.hidden_size,
+            patch_size=self.patch_size,
+            sr_scale=sr_scale,
+            latent_spatial_down_factor=latent_spatial_down_factor,
+            num_res_blocks=lq_num_res_blocks,
+            num_outputs=num_lq_outputs,
+            interval=lq_interval,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts):
+        return precompute_freqs_cis_2d(
+            self.hidden_size // self.num_groups,
+            height, width,
+            ref_grid_h=self.rope_ref_grid_h, ref_grid_w=self.rope_ref_grid_w,
+            device=device, dtype=dtype, **rope_opts,
+        )
+
+    def _pre_patch_block(self, s, i, pid_lq_features, pid_degrade_sigma, **kwargs):
+        if not self.lq_proj.is_gate_active(i):
+            return s
+        out_idx = self.lq_proj.output_index(i)
+        if out_idx >= len(pid_lq_features):
+            return s
+        return self.lq_proj.gate(s, pid_lq_features[out_idx], pid_degrade_sigma, out_idx)
+
+    def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, lq_latent=None, degrade_sigma=None, **kwargs):
+        if lq_latent is None:
+            raise ValueError("PidNet requires lq_latent — attach via PiDConditioning")
+        expected_c = self.lq_proj.latent_channels
+        if lq_latent.shape[1] != expected_c:
+            raise ValueError(
+                f"Input latent has {lq_latent.shape[1]} channels, this model variant expects {expected_c}. "
+                f"Flux1/SD3 = 16 channels, Flux2 = 128 channels."
+            )
+        B = x.shape[0]
+        # Match the backbone's pad_to_patch_size (round up) so the LQ grid lines up with the patch stream.
+        Hs = -(-x.shape[2] // self.patch_size)
+        Ws = -(-x.shape[3] // self.patch_size)
+
+        degrade_sigma = degrade_sigma.to(device=x.device, dtype=torch.float32).reshape(-1)
+        if degrade_sigma.numel() == 1 and B > 1:
+            degrade_sigma = degrade_sigma.expand(B).contiguous()
+
+        lq_features = self.lq_proj(lq_latent=lq_latent.to(x), target_pH=Hs, target_pW=Ws)
+
+        return super()._forward(
+            x, timesteps,
+            context=context, attention_mask=attention_mask,
+            transformer_options=transformer_options,
+            pid_lq_features=lq_features,
+            pid_degrade_sigma=degrade_sigma,
+            **kwargs,
+        )

comfy/ldm/qwen_image/model.py

@@ -51,15 +51,6 @@ class FeedForward(nn.Module):
         return hidden_states
 
 
-def apply_rotary_emb(x, freqs_cis):
-    if x.shape[1] == 0:
-        return x
-
-    t_ = x.reshape(*x.shape[:-1], -1, 1, 2)
-    t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
-    return t_out.reshape(*x.shape)
-
-
 class QwenTimestepProjEmbeddings(nn.Module):
     def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None):
         super().__init__()

comfy/memory_management.py

@@ -4,6 +4,7 @@ import dataclasses
 import torch
 from typing import NamedTuple
 
+import comfy_aimdo.host_buffer
 from comfy.quant_ops import QuantizedTensor
 
 
@@ -17,21 +18,18 @@ class TensorFileSlice(NamedTuple):
 def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=None):
 
     if isinstance(tensor, QuantizedTensor):
-        if not isinstance(destination, QuantizedTensor):
-            return False
-        if tensor._layout_cls != destination._layout_cls:
-            return False
-
-        if not read_tensor_file_slice_into(tensor._qdata, destination._qdata, stream=stream,
+        if not read_tensor_file_slice_into(tensor._qdata,
+                                           destination._qdata if destination is not None else None, stream=stream,
                                            destination2=(destination2._qdata if destination2 is not None else None)):
             return False
 
-        dst_orig_dtype = destination._params.orig_dtype
-        destination._params.copy_from(tensor._params, non_blocking=False)
-        destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype)
+        if destination is not None:
+            dst_orig_dtype = destination._params.orig_dtype
+            destination._params.copy_from(tensor._params, non_blocking=False)
+            destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype)
         if destination2 is not None:
             dst_orig_dtype = destination2._params.orig_dtype
-            destination2._params.copy_from(destination._params, non_blocking=True)
+            destination2._params.copy_from(destination._params if destination is not None else tensor._params, non_blocking=True)
             destination2._params = dataclasses.replace(destination2._params, orig_dtype=dst_orig_dtype)
         return True
 
@@ -39,10 +37,15 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N
     if info is None:
         return False
 
+    if destination is not None and destination.device.type != "cpu" and destination2 is None:
+        destination2 = destination
+        destination = None
+
     file_obj = info.file_ref
-    if (destination.device.type != "cpu"
-            or file_obj is None
-            or destination.numel() * destination.element_size() < info.size
+    if (file_obj is None
+            or (destination is None and destination2 is None)
+            or (destination is not None and (destination.device.type != "cpu" or destination.numel() * destination.element_size() < info.size))
+            or (destination2 is not None and (destination2.device.type == "cpu" or destination2.numel() * destination2.element_size() < info.size))
             or tensor.numel() * tensor.element_size() != info.size
             or tensor.storage_offset() != 0
             or not tensor.is_contiguous()):
@@ -51,6 +54,14 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N
     if info.size == 0:
         return True
 
+    if destination is None:
+        stream_ptr = getattr(stream, "cuda_stream", 0) if stream is not None else 0
+        comfy_aimdo.host_buffer.read_file_to_device(file_obj, info.offset, info.size,
+                                                    stream_ptr, destination2.data_ptr(),
+                                                    destination2.device.index,
+                                                    mark_cold=False)
+        return True
+
     hostbuf = getattr(destination.untyped_storage(), "_comfy_hostbuf", None)
     if hostbuf is not None:
         stream_ptr = getattr(stream, "cuda_stream", 0) if stream is not None else 0
@@ -63,6 +74,9 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N
                                     device=None if destination2 is None else destination2.device.index)
         return True
 
+    if not hasattr(file_obj, "seek") or not hasattr(file_obj, "readinto"):
+        return False
+
     buf_type = ctypes.c_ubyte * info.size
     view = memoryview(buf_type.from_address(destination.data_ptr()))
 

comfy/model_base.py

@@ -50,6 +50,8 @@ import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
 import comfy.ldm.chroma.model
 import comfy.ldm.chroma_radiance.model
+import comfy.ldm.pixeldit.model
+import comfy.ldm.pixeldit.pid
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
@@ -1519,6 +1521,53 @@ class ZImagePixelSpace(Lumina2):
         BaseModel.__init__(self, model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiTPixelSpace)
         self.memory_usage_factor_conds = ("ref_latents",)
 
+
+class PixelDiTT2I(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device,
+                         unet_model=comfy.ldm.pixeldit.model.PixDiT_T2I)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out["attention_mask"] = comfy.conds.CONDRegular(attention_mask)
+        return out
+
+
+class PiD(PixelDiTT2I):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        BaseModel.__init__(self, model_config, model_type, device=device,
+                           unet_model=comfy.ldm.pixeldit.pid.PidNet)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        lq_latent = kwargs.get("lq_latent", None)
+        if lq_latent is not None:
+            out["lq_latent"] = comfy.conds.CONDRegular(lq_latent)
+        degrade_sigma = kwargs.get("degrade_sigma", None)
+        if degrade_sigma is not None:
+            out["degrade_sigma"] = comfy.conds.CONDRegular(degrade_sigma)
+        return out
+
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "lq_latent" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+            lq = cond_value.cond
+            dim = window.dim
+            if dim >= lq.ndim:
+                return None
+            lq_proj = self.diffusion_model.lq_proj
+            ratio = lq_proj.sr_scale * lq_proj.latent_spatial_down_factor
+            # Map x window indices -> lq indices (deduplicated, sorted, in-bounds).
+            lq_size = lq.size(dim)
+            lq_indices = sorted({i // ratio for i in window.index_list if 0 <= i // ratio < lq_size})
+            if not lq_indices:
+                return None
+            idx = tuple([slice(None)] * dim + [lq_indices])
+            return cond_value._copy_with(lq[idx].to(device))
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
+
 class WAN21(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)

comfy/model_detection.py

@@ -463,6 +463,23 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["extra_per_block_abs_pos_emb_type"] = "learnable"
         return dit_config
 
+    # PiD (Pixel Diffusion Decoder). Must check BEFORE plain PixelDiT_T2I.
+    _lq_w_key = '{}lq_proj.latent_proj.0.weight'.format(key_prefix)
+    if _lq_w_key in state_dict_keys:
+        in_ch = int(state_dict[_lq_w_key].shape[1])
+        _gate_prefix = '{}lq_proj.gate_modules.'.format(key_prefix)
+        num_gates = len({k[len(_gate_prefix):].split('.')[0]
+                         for k in state_dict_keys if k.startswith(_gate_prefix)})
+        dit_config = {"image_model": "pid",
+                      "lq_latent_channels": in_ch,
+                      "latent_spatial_down_factor": 16 if in_ch >= 64 else 8}
+        if num_gates > 0:
+            dit_config["lq_interval"] = (14 + num_gates - 1) // num_gates
+        return dit_config
+
+    if '{}core.pixel_embedder.proj.weight'.format(key_prefix) in state_dict_keys:  # PixelDiT T2I
+        return {"image_model": "pixeldit_t2i"}
+
     if '{}cap_embedder.1.weight'.format(key_prefix) in state_dict_keys and '{}noise_refiner.0.attention.k_norm.weight'.format(key_prefix) in state_dict_keys:  # Lumina 2
         dit_config = {}
         dit_config["image_model"] = "lumina2"

comfy/model_management.py

@@ -641,14 +641,17 @@ def free_pins(size, evict_active=False):
     return freed_total
 
 def ensure_pin_budget(size, evict_active=False):
-    shortfall = size + comfy.memory_management.RAM_CACHE_HEADROOM / 2 - psutil.virtual_memory().available
+    if args.fast_disk:
+        shortfall = TOTAL_PINNED_MEMORY + size - MAX_PINNED_MEMORY
+    else:
+        shortfall = size + max(comfy.memory_management.RAM_CACHE_HEADROOM / 2, 2048 * 1024 ** 2) - psutil.virtual_memory().available
     if shortfall <= 0:
         return True
 
     to_free = shortfall + PIN_PRESSURE_HYSTERESIS
     return free_pins(to_free, evict_active=evict_active) >= shortfall
 
-def ensure_pin_registerable(size, evict_active=False):
+def ensure_pin_registerable(size, evict_active=True):
     shortfall = TOTAL_PINNED_MEMORY + size - MAX_PINNED_MEMORY
     if MAX_PINNED_MEMORY <= 0:
         return False
@@ -658,10 +661,17 @@ def ensure_pin_registerable(size, evict_active=False):
     shortfall += REGISTERABLE_PIN_HYSTERESIS
     for loaded_model in reversed(current_loaded_models):
         model = loaded_model.model
-        if model is not None and model.is_dynamic() and (evict_active or not model.model.dynamic_pins[model.load_device]["active"]):
+        if model is not None and model.is_dynamic() and not model.model.dynamic_pins[model.load_device]["active"]:
             shortfall -= model.unregister_inactive_pins(shortfall)
             if shortfall <= 0:
                 return True
+    if evict_active:
+        for loaded_model in current_loaded_models:
+            model = loaded_model.model
+            if model is not None and model.is_dynamic() and model.model.dynamic_pins[model.load_device]["active"]:
+                shortfall -= model.unregister_inactive_pins(shortfall)
+                if shortfall <= 0:
+                    return True
     return shortfall <= REGISTERABLE_PIN_HYSTERESIS
 
 class LoadedModel:
@@ -803,9 +813,9 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins
     for x in can_unload_sorted:
         i = x[-1]
         memory_to_free = 1e32
-        if current_loaded_models[i].model.is_dynamic() and (not DISABLE_SMART_MEMORY or device is None):
+        if not DISABLE_SMART_MEMORY or device is None:
             memory_to_free = 0 if device is None else memory_required - get_free_memory(device)
-            if for_dynamic:
+            if current_loaded_models[i].model.is_dynamic() and for_dynamic:
                 #don't actually unload dynamic models for the sake of other dynamic models
                 #as that works on-demand.
                 memory_required -= current_loaded_models[i].model.loaded_size()
@@ -817,6 +827,10 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins
     for i in sorted(unloaded_model, reverse=True):
         unloaded_models.append(current_loaded_models.pop(i))
 
+    if not for_dynamic and pins_required > 0:
+        ensure_pin_budget(pins_required)
+        ensure_pin_registerable(pins_required)
+
     if len(unloaded_model) > 0:
         soft_empty_cache()
     elif device is not None:
@@ -879,15 +893,19 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             model_to_unload.model_finalizer.detach()
 
     total_memory_required = {}
+    total_pins_required = {}
     for loaded_model in models_to_load:
         device = loaded_model.device
         total_memory_required[device] = total_memory_required.get(device, 0) + loaded_model.model_memory_required(device)
+        if not loaded_model.model.is_dynamic():
+            total_pins_required[device] = total_pins_required.get(device, 0) + loaded_model.model_memory()
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
             free_memory(total_memory_required[device] * 1.1 + extra_mem,
                         device,
-                        for_dynamic=free_for_dynamic)
+                        for_dynamic=free_for_dynamic,
+                        pins_required=total_pins_required.get(device, 0))
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
@@ -1283,7 +1301,6 @@ STREAM_CAST_BUFFERS = {}
 LARGEST_CASTED_WEIGHT = (None, 0)
 STREAM_AIMDO_CAST_BUFFERS = {}
 LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
-STREAM_PIN_BUFFERS = {}
 
 DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3
 
@@ -1326,42 +1343,13 @@ def get_aimdo_cast_buffer(offload_stream, device):
         STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer
     return cast_buffer
 
-def get_pin_buffer(offload_stream):
-    pin_buffer = STREAM_PIN_BUFFERS.get(offload_stream, None)
-    if pin_buffer is None:
-        pin_buffer = comfy_aimdo.host_buffer.HostBuffer(0, 0, pinned_hostbuf_size(8 * 1024**3), mark_cold=False)
-        STREAM_PIN_BUFFERS[offload_stream] = pin_buffer
-    elif offload_stream is not None:
-        event = getattr(pin_buffer, "_comfy_event", None)
-        if event is not None:
-            event.synchronize()
-            delattr(pin_buffer, "_comfy_event")
-    return pin_buffer
-
-def resize_pin_buffer(pin_buffer, size):
-    global TOTAL_PINNED_MEMORY
-    old_size = pin_buffer.size
-    if size <= old_size:
-        return True
-    growth = size - old_size
-    comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM)
-    ensure_pin_budget(growth, evict_active=True)
-    ensure_pin_registerable(growth, evict_active=True)
-    try:
-        pin_buffer.extend(size=size, reallocate=True)
-    except RuntimeError:
-        return False
-    TOTAL_PINNED_MEMORY += pin_buffer.size - old_size
-    return True
-
 def reset_cast_buffers():
-    global TOTAL_PINNED_MEMORY
     global LARGEST_CASTED_WEIGHT
     global LARGEST_AIMDO_CASTED_WEIGHT
 
     LARGEST_CASTED_WEIGHT = (None, 0)
     LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
-    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS) | set(STREAM_PIN_BUFFERS):
+    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS):
         if offload_stream is not None:
             offload_stream.synchronize()
     synchronize()
@@ -1370,20 +1358,24 @@ def reset_cast_buffers():
         mmap_obj.bounce()
     DIRTY_MMAPS.clear()
 
-    for pin_buffer in STREAM_PIN_BUFFERS.values():
-        TOTAL_PINNED_MEMORY -= pin_buffer.size
-    TOTAL_PINNED_MEMORY = max(0, TOTAL_PINNED_MEMORY)
-
     for loaded_model in current_loaded_models:
         model = loaded_model.model
         if model is not None and model.is_dynamic():
-            model.model.dynamic_pins[model.load_device]["active"] = False
+            pin_state = model.model.dynamic_pins[model.load_device]
+
+            if pin_state["active"]:
+                *_, buckets = pin_state["weights"]
+                for size, bucket in list(buckets.items()):
+                    bucket[:] = [ entry for entry in bucket if entry[-1] is not None ]
+                    if not bucket:
+                        del buckets[size]
+
+            pin_state["active"] = False
             model.partially_unload_ram(1e30, subsets=[ "patches" ])
-            model.model.dynamic_pins[model.load_device]["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, pinned_hostbuf_size(model.model_size())), [], [-1], [0])
+            model.model.dynamic_pins[model.load_device]["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, pinned_hostbuf_size(model.model_size())), [], [-1], [0], [0], {})
 
     STREAM_CAST_BUFFERS.clear()
     STREAM_AIMDO_CAST_BUFFERS.clear()
-    STREAM_PIN_BUFFERS.clear()
     soft_empty_cache()
 
 def get_offload_stream(device):
@@ -1436,7 +1428,7 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None, r2=None):
        if hasattr(wf_context, "as_context"):
            wf_context = wf_context.as_context(stream)
 
-    dest_views = comfy.memory_management.interpret_gathered_like(tensors, r)
+    dest_views = comfy.memory_management.interpret_gathered_like(tensors, r) if r is not None else [None] * len(tensors)
     dest2_views = comfy.memory_management.interpret_gathered_like(tensors, r2) if r2 is not None else None
     with wf_context:
         for tensor in tensors:
@@ -1448,9 +1440,10 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None, r2=None):
                 continue
             storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage()
             mark_mmap_dirty(storage)
-            dest_view.copy_(tensor, non_blocking=non_blocking)
+            if dest_view is not None:
+                dest_view.copy_(tensor, non_blocking=non_blocking)
             if dest2_view is not None:
-                dest2_view.copy_(dest_view, non_blocking=non_blocking)
+                dest2_view.copy_(tensor if dest_view is None else dest_view, non_blocking=non_blocking)
 
 
 def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None, r=None):
@@ -1723,6 +1716,13 @@ def is_device_xpu(device):
 def is_device_cuda(device):
     return is_device_type(device, 'cuda')
 
+def set_torch_device(device):
+    """Set the current device for the given torch device. Supports CUDA and XPU."""
+    if is_device_cuda(device):
+        torch.cuda.set_device(device)
+    elif is_device_xpu(device):
+        torch.xpu.set_device(device)
+
 def is_directml_enabled():
     global directml_enabled
     if directml_enabled:

comfy/model_patcher.py

@@ -1721,8 +1721,8 @@ class ModelPatcherDynamic(ModelPatcher):
         """
         if device not in self.model.dynamic_pins:
             self.model.dynamic_pins[device] = {
-                "weights": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]),
-                "patches": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]),
+                "weights": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0], [0], {}),
+                "patches": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0], [0], {}),
                 "hostbufs_initialized": False,
                 "failed": False,
                 "active": False,
@@ -1799,8 +1799,8 @@ class ModelPatcherDynamic(ModelPatcher):
             pin_state = self.model.dynamic_pins[self.load_device]
             if not pin_state["hostbufs_initialized"]:
                 hostbuf_size = comfy.model_management.pinned_hostbuf_size(self.model_size())
-                pin_state["weights"] = (comfy_aimdo.host_buffer.HostBuffer(0, 64 * 1024 * 1024, hostbuf_size), [], [-1], [0])
-                pin_state["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, hostbuf_size), [], [-1], [0])
+                pin_state["weights"] = (comfy_aimdo.host_buffer.HostBuffer(0, 64 * 1024 * 1024, hostbuf_size), [], [-1], [0], [0], {})
+                pin_state["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, hostbuf_size), [], [-1], [0], [0], {})
                 pin_state["hostbufs_initialized"] = True
             pin_state["failed"] = False
             pin_state["active"] = True
@@ -1942,18 +1942,16 @@ class ModelPatcherDynamic(ModelPatcher):
         return freed
 
     def loaded_ram_size(self):
-        return (self.model.dynamic_pins[self.load_device]["weights"][0].size +
-                self.model.dynamic_pins[self.load_device]["patches"][0].size)
+        return (self.model.dynamic_pins[self.load_device]["weights"][0].size)
 
     def pinned_memory_size(self):
-        return (self.model.dynamic_pins[self.load_device]["weights"][3][0] +
-                self.model.dynamic_pins[self.load_device]["patches"][3][0])
+        return (self.model.dynamic_pins[self.load_device]["weights"][3][0])
 
     def unregister_inactive_pins(self, ram_to_unload, subsets=[ "weights", "patches" ]):
         freed = 0
         pin_state = self.model.dynamic_pins[self.load_device]
         for subset in subsets:
-            hostbuf, stack, stack_split, pinned_size = pin_state[subset]
+            hostbuf, stack, stack_split, pinned_size, *_ = pin_state[subset]
             split = stack_split[0]
             while split >= 0:
                 module, offset = stack[split]
@@ -1978,10 +1976,12 @@ class ModelPatcherDynamic(ModelPatcher):
         freed = 0
         pin_state = self.model.dynamic_pins[self.load_device]
         for subset in subsets:
-            hostbuf, stack, stack_split, pinned_size = pin_state[subset]
+            hostbuf, stack, stack_split, pinned_size, *_ = pin_state[subset]
             while len(stack) > 0:
                 module, offset = stack.pop()
                 size = module._pin.numel() * module._pin.element_size()
+                module._pin_balancer_entry[-1] = None
+                del module._pin_balancer_entry
                 del module._pin
                 hostbuf.truncate(offset, do_unregister=module._pin_registered)
                 stack_split[0] = min(stack_split[0], len(stack) - 1)

comfy/model_prefetch.py

@@ -1,4 +1,5 @@
 import comfy_aimdo.model_vbar
+import comfy.memory_management
 import comfy.model_management
 import comfy.ops
 
@@ -50,7 +51,17 @@ def prefetch_queue_pop(queue, device, module):
             if hasattr(s, "_v"):
                 comfy_modules.append(s)
 
+        registerable_size = 0
+        for s in comfy_modules:
+            registerable_size += comfy.memory_management.vram_aligned_size([s.weight, s.bias])
+            for param_key in ("weight", "bias"):
+                lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+                if lowvram_fn is not None:
+                    registerable_size += lowvram_fn.memory_required()
+
         offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True)
+        if not comfy.model_management.args.fast_disk:
+            comfy.model_management.ensure_pin_registerable(registerable_size)
         comfy.model_management.sync_stream(device, offload_stream)
         queue[0] = (offload_stream, (prefetch, comfy_modules))
 

comfy/multigpu.py

@@ -17,7 +17,7 @@ class MultiGPUThreadPool:
     """Persistent thread pool for multi-GPU work distribution.
 
     Maintains one worker thread per extra GPU device. Each thread calls
-    torch.cuda.set_device() once at startup so that compiled kernel caches
+    set_torch_device() once at startup so that compiled kernel caches
     (inductor/triton) stay warm across diffusion steps.
     """
 
@@ -37,7 +37,7 @@ class MultiGPUThreadPool:
 
     def _worker_loop(self, device: torch.device, work_q: queue.Queue, result_q: queue.Queue):
         try:
-            torch.cuda.set_device(device)
+            comfy.model_management.set_torch_device(device)
         except Exception as e:
             logging.error(f"MultiGPUThreadPool: failed to set device {device}: {e}")
             while True:

comfy/ops.py

@@ -76,8 +76,6 @@ except:
 
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
-STREAM_PIN_BUFFER_HEADROOM = 8 * 1024 * 1024
-
 def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
@@ -94,9 +92,6 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin
     offload_stream = None
     cast_buffer = None
     cast_buffer_offset = 0
-    stream_pin_hostbuf = None
-    stream_pin_offset = 0
-    stream_pin_queue = []
 
     def ensure_offload_stream(module, required_size, check_largest):
         nonlocal offload_stream
@@ -130,22 +125,6 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin
         cast_buffer_offset += buffer_size
         return buffer
 
-    def get_stream_pin_buffer_offset(buffer_size):
-        nonlocal stream_pin_hostbuf
-        nonlocal stream_pin_offset
-
-        if buffer_size == 0 or offload_stream is None:
-            return None
-
-        if stream_pin_hostbuf is None:
-            stream_pin_hostbuf = comfy.model_management.get_pin_buffer(offload_stream)
-            if stream_pin_hostbuf is None:
-                return None
-
-        offset = stream_pin_offset
-        stream_pin_offset += buffer_size
-        return offset
-
     for s in comfy_modules:
         signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
         resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
@@ -184,12 +163,18 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin
         if xfer_dest is None:
             xfer_dest = get_cast_buffer(dest_size)
 
-        def cast_maybe_lowvram_patch(xfer_source, xfer_dest, stream):
+        def cast_maybe_lowvram_patch(xfer_source, xfer_dest, stream, xfer_dest2=None):
             if xfer_source is not None:
                 if getattr(xfer_source, "is_lowvram_patch", False):
-                    xfer_source.prepare(xfer_dest, stream, copy=True, commit=False)
-                else:
-                    comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=stream)
+                    if xfer_dest is not None:
+                        xfer_source.prepare(xfer_dest, stream, copy=True, commit=False)
+                        xfer_source = [ xfer_dest ]
+                        xfer_dest = xfer_dest2
+                        xfer_dest2 = None
+                    elif xfer_dest2 is not None:
+                        xfer_source.prepare(xfer_dest2, stream, copy=True, commit=False)
+                        return
+                comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=stream, r2=xfer_dest2)
 
         def handle_pin(m, pin, source, dest, subset="weights", size=None):
             if pin is not None:
@@ -198,19 +183,7 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin
             if signature is None:
                 comfy.pinned_memory.pin_memory(m, subset=subset, size=size)
                 pin = comfy.pinned_memory.get_pin(m, subset=subset)
-                if pin is not None:
-                    if isinstance(source, list):
-                        comfy.model_management.cast_to_gathered(source, pin, non_blocking=non_blocking, stream=offload_stream, r2=dest)
-                    else:
-                        cast_maybe_lowvram_patch(source, pin, None)
-                        cast_maybe_lowvram_patch([ pin ], dest, offload_stream)
-                    return
-            if pin is None:
-                pin_offset = get_stream_pin_buffer_offset(size)
-                if pin_offset is not None:
-                    stream_pin_queue.append((source, pin_offset, size, dest))
-                    return
-            cast_maybe_lowvram_patch(source, dest, offload_stream)
+            cast_maybe_lowvram_patch(source, pin, offload_stream, xfer_dest2=dest)
 
         handle_pin(s, pin, xfer_source, xfer_dest, size=dest_size)
 
@@ -232,23 +205,6 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin
         prefetch["needs_cast"] = needs_cast
         s._prefetch = prefetch
 
-    if stream_pin_offset > 0:
-        if stream_pin_hostbuf.size < stream_pin_offset:
-            if not comfy.model_management.resize_pin_buffer(stream_pin_hostbuf, stream_pin_offset + STREAM_PIN_BUFFER_HEADROOM):
-                for xfer_source, _, _, xfer_dest in stream_pin_queue:
-                    cast_maybe_lowvram_patch(xfer_source, xfer_dest, offload_stream)
-                return offload_stream
-        stream_pin_tensor = comfy_aimdo.torch.hostbuf_to_tensor(stream_pin_hostbuf)
-        stream_pin_tensor.untyped_storage()._comfy_hostbuf = stream_pin_hostbuf
-        for xfer_source, pin_offset, pin_size, xfer_dest in stream_pin_queue:
-            pin = stream_pin_tensor[pin_offset:pin_offset + pin_size]
-            if isinstance(xfer_source, list):
-                comfy.model_management.cast_to_gathered(xfer_source, pin, non_blocking=non_blocking, stream=offload_stream, r2=xfer_dest)
-            else:
-                cast_maybe_lowvram_patch(xfer_source, pin, None)
-                comfy.model_management.cast_to_gathered([ pin ], xfer_dest, non_blocking=non_blocking, stream=offload_stream)
-        stream_pin_hostbuf._comfy_event = offload_stream.record_event()
-
     return offload_stream
 
 

comfy/pinned_memory.py

@@ -1,17 +1,55 @@
+import bisect
+
 import comfy.model_management
 import comfy.memory_management
+import comfy.utils
 import comfy_aimdo.host_buffer
 import comfy_aimdo.torch
 import torch
 
 from comfy.cli_args import args
 
+def _add_to_bucket(module, buckets, size, priority):
+    bucket = buckets.setdefault(size, [])
+    entry = [-priority, 0, module]
+    entry[1] = id(entry)
+    bisect.insort(bucket, entry)
+    module._pin_balancer_entry = entry
+
+def _steal_pin(module, stack, buckets, size, priority):
+    bucket = buckets.get(size)
+    if bucket is None:
+        return False
+
+    while bucket and bucket[-1][-1] is None:
+        bucket.pop()
+    if not bucket:
+        del buckets[size]
+        return False
+
+    if priority <= -bucket[-1][0]:
+        return False
+
+    *_, victim = bucket.pop()
+    module._pin = victim._pin
+    module._pin_registered = victim._pin_registered
+    module._pin_stack_index = victim._pin_stack_index
+    stack[module._pin_stack_index] = (module, stack[module._pin_stack_index][1])
+
+    victim._pin_registered = False
+    del victim._pin
+    del victim._pin_stack_index
+    del victim._pin_balancer_entry
+
+    _add_to_bucket(module, buckets, size, priority)
+    return True
+
 def get_pin(module, subset="weights"):
     pin = getattr(module, "_pin", None)
     if pin is None or module._pin_registered or args.disable_pinned_memory:
         return pin
 
-    _, _, stack_split, pinned_size = module._pin_state[subset]
+    _, _, stack_split, pinned_size, *_ = module._pin_state[subset]
     size = pin.nbytes
     comfy.model_management.ensure_pin_registerable(size)
 
@@ -31,26 +69,30 @@ def pin_memory(module, subset="weights", size=None):
         return
 
     pin = get_pin(module, subset)
-    if pin is not None or pin_state["failed"]:
+    if pin is not None:
         return
 
-    hostbuf, stack, stack_split, pinned_size = pin_state[subset]
+    hostbuf, stack, stack_split, pinned_size, counter, buckets = pin_state[subset]
     if size is None:
         size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
     offset = hostbuf.size
-    registerable_size = size + max(0, hostbuf.size - pinned_size[0])
+    registerable_size = size
+    priority = getattr(module, "_pin_balancer_priority", None)
+
+    if priority is None:
+        priority = comfy.utils.bit_reverse_range(counter[0], 16)
+        counter[0] += 1
+        module._pin_balancer_priority = priority
 
     comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM)
     if (not comfy.model_management.ensure_pin_budget(size) or
         not comfy.model_management.ensure_pin_registerable(registerable_size)):
-        pin_state["failed"] = True
-        return False
+        return _steal_pin(module, stack, buckets, size, priority)
 
     try:
         hostbuf.extend(size=size)
     except RuntimeError:
-        pin_state["failed"] = True
-        return False
+        return _steal_pin(module, stack, buckets, size, priority)
 
     module._pin = comfy_aimdo.torch.hostbuf_to_tensor(hostbuf)[offset:offset + size]
     module._pin.untyped_storage()._comfy_hostbuf = hostbuf
@@ -60,4 +102,5 @@ def pin_memory(module, subset="weights", size=None):
     stack_split[0] = max(stack_split[0], module._pin_stack_index)
     comfy.model_management.TOTAL_PINNED_MEMORY += size
     pinned_size[0] += size
+    _add_to_bucket(module, buckets, size, priority)
     return True

comfy/samplers.py

@@ -464,10 +464,7 @@ def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: t
 
     def _handle_batch(device: torch.device, batch_tuple: tuple[comfy.hooks.HookGroup, tuple], results: list[thread_result]):
         try:
-            # TODO: non-NVIDIA support -- guard with `if device.type == "cuda":` once
-            # we extend multigpu QA beyond CUDA. Unconditional call crashes on
-            # XPU/NPU/MPS/CPU/DirectML backends.
-            torch.cuda.set_device(device)
+            comfy.model_management.set_torch_device(device)
             model_current: BaseModel = model_options["multigpu_clones"][device].model
             # run every hooked_to_run separately
             with torch.no_grad():

comfy/sd.py

@@ -49,6 +49,7 @@ import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
+import comfy.text_encoders.pixeldit
 import comfy.text_encoders.wan
 import comfy.text_encoders.hidream
 import comfy.text_encoders.ace
@@ -1285,6 +1286,7 @@ class CLIPType(Enum):
     LONGCAT_IMAGE = 26
     COGVIDEOX = 27
     LENS = 28
+    PIXELDIT = 29
 
 
 
@@ -1528,8 +1530,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.tokenizer = variant.tokenizer
             tokenizer_data["tokenizer_json"] = clip_data[0].get("tokenizer_json", None)
         elif te_model == TEModel.GEMMA_2_2B:
-            clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
+            if clip_type == CLIPType.PIXELDIT:
+                clip_target.clip = comfy.text_encoders.pixeldit.pixeldit_te(**llama_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.pixeldit.PixelDiTGemma2Tokenizer
+            else:
+                clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
             tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
         elif te_model == TEModel.GEMMA_3_4B:
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data), model_type="gemma3_4b")

comfy/supported_models.py

@@ -30,6 +30,7 @@ import comfy.text_encoders.longcat_image
 import comfy.text_encoders.ernie
 import comfy.text_encoders.cogvideo
 import comfy.text_encoders.hidream_o1
+import comfy.text_encoders.pixeldit
 
 from . import supported_models_base
 from . import latent_formats
@@ -844,6 +845,8 @@ class Lens(supported_models_base.BASE):
     unet_extra_config = {}
     latent_format = latent_formats.Flux2
 
+    memory_usage_factor = 4.0
+
     supported_inference_dtypes = [torch.bfloat16, torch.float32] # fp16 causes NaNs
 
     vae_key_prefix = ["vae."]
@@ -1201,6 +1204,72 @@ class ZImagePixelSpace(ZImage):
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.ZImagePixelSpace(self, device=device)
 
+class PixelDiTT2I(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "pixeldit_t2i",
+    }
+
+    unet_extra_config = {}
+
+    sampling_settings = {
+        "shift": 4.0,  # 1024px stage 3 default; 2.0 for 512px
+    }
+
+    latent_format = latent_formats.PixelDiTPixel
+    memory_usage_factor = 0.04
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.PixelDiTT2I(self, device=device)
+
+    def process_unet_state_dict(self, state_dict):
+        # pixel_dim from pixel_embedder.proj.weight = (pixel_dim, in_channels); p2 derived per-weight from total // (6 * pixel_dim).
+        pixel_dim = next(v for k, v in state_dict.items() if k.endswith("pixel_embedder.proj.weight")).shape[0]
+
+        out = {}
+        marker = ".adaLN_modulation.0."
+        for k, v in state_dict.items():
+            if k.startswith("_repa_projector") or k.startswith("net_ema."):
+                continue
+            if k.startswith("core."):
+                k = k[len("core."):]
+            elif k.startswith("net."):
+                k = k[len("net."):]
+            if "pixel_blocks." in k and marker in k:
+                # Split into msa (chunks 0-2) and mlp (chunks 3-5) for the two-Linear PiTBlock to reduce peak VRAM
+                p2 = v.shape[0] // (6 * pixel_dim)
+                trail = v.shape[1:]  # () for bias, (in_dim,) for weight
+                vv = v.view(p2, 6, pixel_dim, *trail)
+                base, suffix = k.split(marker)
+                out[f"{base}.adaLN_modulation_msa.{suffix}"] = vv[:, 0:3].reshape(3 * p2 * pixel_dim, *trail).contiguous()
+                out[f"{base}.adaLN_modulation_mlp.{suffix}"] = vv[:, 3:6].reshape(3 * p2 * pixel_dim, *trail).contiguous()
+            else:
+                out[k] = v
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(
+            comfy.text_encoders.pixeldit.PixelDiTGemma2Tokenizer,
+            comfy.text_encoders.pixeldit.PixelDiTGemma2TE,
+        )
+
+class PiD(PixelDiTT2I):
+    unet_config = {
+        "image_model": "pid",
+    }
+
+    sampling_settings = {
+        "shift": 1.5, # close approximation of the original distill 4 steps [0.999, 0.866, 0.634, 0.342, 0]
+    }
+
+    memory_usage_factor = 0.04
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.PiD(self, device=device)
+
 class WAN21_T2V(supported_models_base.BASE):
     unet_config = {
         "image_model": "wan2.1",
@@ -2111,6 +2180,8 @@ models = [
     CosmosI2VPredict2,
     ZImagePixelSpace,
     ZImage,
+    PiD,
+    PixelDiTT2I,
     Lumina2,
     WAN22_T2V,
     WAN21_CausalAR_T2V,

comfy/text_encoders/pixeldit.py

@@ -0,0 +1,104 @@
+import torch
+
+from comfy import sd1_clip
+from .lumina2 import Gemma2BTokenizer, LuminaModel
+import comfy.text_encoders.llama
+
+
+class PixelDiTGemma2_2BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(
+            device=device, layer=layer, layer_idx=layer_idx,
+            textmodel_json_config={}, dtype=dtype,
+            special_tokens={"start": 2, "pad": 0},
+            layer_norm_hidden_state=False,
+            model_class=comfy.text_encoders.llama.Gemma2_2B,
+            enable_attention_masks=attention_mask,
+            return_attention_masks=attention_mask,
+            model_options=model_options,
+        )
+
+
+_PIXELDIT_CHI_PROMPT = (
+    'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual descriptions '
+    "suitable for image generation. Evaluate the level of detail in the user prompt:\n"
+    "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, "
+    "and spatial relationships to create vivid and concrete scenes.\n"
+    "- If the prompt is already detailed, refine and enhance the existing details slightly without "
+    "overcomplicating.\n"
+    "Here are examples of how to transform or refine prompts:\n"
+    "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, "
+    "sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.\n"
+    "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring "
+    "glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus "
+    "passing by towering glass skyscrapers.\n"
+    "Please generate only the enhanced description for the prompt below and avoid including any "
+    "additional commentary or evaluations:\n"
+    "User Prompt: "
+)
+
+_PIXELDIT_MAX_LENGTH = 300
+_PIXELDIT_CHI_PROMPT_DETECT_PREFIX = 'Given a user prompt, generate an "Enhanced prompt"'
+
+
+class PixelDiTGemma2Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data=None):
+        if tokenizer_data is None:
+            tokenizer_data = {}
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data,
+                         name="gemma2_2b", tokenizer=Gemma2BTokenizer)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+        if not text.strip():
+            return super().tokenize_with_weights("", return_word_ids=return_word_ids, disable_weights=True, min_length=_PIXELDIT_MAX_LENGTH)
+
+        chi_token_count = len(self.gemma2_2b.tokenizer(_PIXELDIT_CHI_PROMPT)["input_ids"])
+        combined = text if text.startswith(_PIXELDIT_CHI_PROMPT_DETECT_PREFIX) else _PIXELDIT_CHI_PROMPT + text
+        max_length_all = chi_token_count + _PIXELDIT_MAX_LENGTH - 2
+        out = super().tokenize_with_weights(combined, return_word_ids=return_word_ids,
+                                            disable_weights=True, min_length=max_length_all)
+        out["gemma2_2b"] = [out["gemma2_2b"][0][:max_length_all]]
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.gemma2_2b.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return self.gemma2_2b.state_dict()
+
+
+class PixelDiTGemma2TE(LuminaModel):
+    # PixelDiT's select_index: keep BOS + last 299 embeddings of the padded sequence.
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="gemma2_2b",
+                         clip_model=PixelDiTGemma2_2BModel, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        result = super().encode_token_weights(token_weight_pairs)
+        cond, pooled = result[0], result[1]
+        extra = result[2] if len(result) > 2 else None
+        if cond.shape[1] > _PIXELDIT_MAX_LENGTH:
+            cond = torch.cat([cond[:, :1], cond[:, -(_PIXELDIT_MAX_LENGTH - 1):]], dim=1)
+            if extra is not None and "attention_mask" in extra:
+                am = extra["attention_mask"]
+                extra["attention_mask"] = torch.cat([am[..., :1], am[..., -(_PIXELDIT_MAX_LENGTH - 1):]], dim=-1)
+        if extra is not None:
+            return cond, pooled, extra
+        return cond, pooled
+
+
+def pixeldit_te(dtype_llama=None, llama_quantization_metadata=None):
+    class PixelDiTTE_(PixelDiTGemma2TE):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return PixelDiTTE_

comfy/utils.py

@@ -85,9 +85,9 @@ _TYPES = {
 def load_safetensors(ckpt):
     import comfy_aimdo.model_mmap
 
-    f = open(ckpt, "rb", buffering=0)
     file_lock = threading.Lock()
     model_mmap = comfy_aimdo.model_mmap.ModelMMAP(ckpt)
+    f = model_mmap.get_file_handle()
     file_size = os.path.getsize(ckpt)
     mv = memoryview((ctypes.c_uint8 * file_size).from_address(model_mmap.get()))
 
@@ -1452,3 +1452,10 @@ def deepcopy_list_dict(obj, memo=None):
 
     memo[obj_id] = res
     return res
+
+def bit_reverse_range(index, bits):
+    result = 0
+    for _ in range(bits):
+        result = (result << 1) | (index & 1)
+        index >>= 1
+    return result

comfy_api/latest/__init__.py

@@ -5,7 +5,7 @@ from comfy_api.internal.singleton import ProxiedSingleton
 from comfy_api.internal.async_to_sync import create_sync_class
 from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
 from ._input_impl import VideoFromFile, VideoFromComponents
-from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, File3D
+from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, SPLAT, File3D
 from . import _io_public as io
 from . import _ui_public as ui
 from comfy_execution.utils import get_executing_context
@@ -143,6 +143,7 @@ class Types:
     VideoComponents = VideoComponents
     MESH = MESH
     VOXEL = VOXEL
+    SPLAT = SPLAT
     File3D = File3D
 
 

comfy_api/latest/_io.py

@@ -28,7 +28,7 @@ if TYPE_CHECKING:
 from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class,
     prune_dict, shallow_clone_class)
 from comfy_execution.graph_utils import ExecutionBlocker
-from ._util import MESH, VOXEL, SVG as _SVG, File3D
+from ._util import MESH, VOXEL, SPLAT, SVG as _SVG, File3D
 
 
 class FolderType(str, Enum):
@@ -684,6 +684,10 @@ class Voxel(ComfyTypeIO):
 class Mesh(ComfyTypeIO):
     Type = MESH
 
+@comfytype(io_type="SPLAT")
+class Splat(ComfyTypeIO):
+    Type = SPLAT
+
 
 @comfytype(io_type="FILE_3D")
 class File3DAny(ComfyTypeIO):
@@ -727,6 +731,30 @@ class File3DUSDZ(ComfyTypeIO):
     Type = File3D
 
 
+@comfytype(io_type="FILE_3D_PLY")
+class File3DPLY(ComfyTypeIO):
+    """PLY format 3D file - point cloud or Gaussian splat."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_SPLAT")
+class File3DSPLAT(ComfyTypeIO):
+    """SPLAT format 3D file - 3D Gaussian splat."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_SPZ")
+class File3DSPZ(ComfyTypeIO):
+    """SPZ format 3D file - compressed 3D Gaussian splat."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_KSPLAT")
+class File3DKSPLAT(ComfyTypeIO):
+    """KSPLAT format 3D file - 3D Gaussian splat."""
+    Type = File3D
+
+
 @comfytype(io_type="HOOKS")
 class Hooks(ComfyTypeIO):
     if TYPE_CHECKING:
@@ -762,14 +790,32 @@ class Accumulation(ComfyTypeIO):
 @comfytype(io_type="LOAD3D_CAMERA")
 class Load3DCamera(ComfyTypeIO):
     class CameraInfo(TypedDict):
-        position: dict[str, float | int]
-        target: dict[str, float | int]
-        zoom: int
-        cameraType: str
+        # Coordinate system: right-handed, Y-up, camera looks down -Z
+        position: dict[str, float | int]  # scene units
+        target: dict[str, float | int]  # scene units; OrbitControls focus point
+        zoom: float | int  # dimensionless, 1 = 100%
+        cameraType: str  # 'perspective' | 'orthographic'
+        quaternion: NotRequired[dict[str, float | int]]  # normalized, dimensionless; camera world rotation
+        fov: NotRequired[float | int]  # degrees, vertical FOV (perspective only)
+        aspect: NotRequired[float | int]  # width / height (perspective only)
+        near: NotRequired[float | int]  # scene units
+        far: NotRequired[float | int]  # scene units
+        frustum: NotRequired[dict[str, float | int]]  # orthographic only: {left, right, top, bottom} in scene units
 
     Type = CameraInfo
 
 
+@comfytype(io_type="LOAD3D_MODEL_INFO")
+class Load3DModelInfo(ComfyTypeIO):
+    class Model3DTransform(TypedDict):
+        # Coordinate system: right-handed, Y-up, world space
+        position: dict[str, float | int]  # scene units
+        quaternion: dict[str, float | int]  # normalized, dimensionless; world rotation
+        scale: dict[str, float | int]  # dimensionless multiplier
+
+    Type = list[Model3DTransform]
+
+
 @comfytype(io_type="LOAD_3D")
 class Load3D(ComfyTypeIO):
     """3D models are stored as a dictionary."""
@@ -779,6 +825,7 @@ class Load3D(ComfyTypeIO):
         normal: str
         camera_info: Load3DCamera.CameraInfo
         recording: NotRequired[str]
+        model_3d_info: NotRequired[list[Load3DModelInfo.Model3DTransform]]
 
     Type = Model3DDict
 
@@ -2277,6 +2324,7 @@ __all__ = [
     "LossMap",
     "Voxel",
     "Mesh",
+    "Splat",
     "File3DAny",
     "File3DGLB",
     "File3DGLTF",
@@ -2284,6 +2332,10 @@ __all__ = [
     "File3DOBJ",
     "File3DSTL",
     "File3DUSDZ",
+    "File3DPLY",
+    "File3DSPLAT",
+    "File3DSPZ",
+    "File3DKSPLAT",
     "Hooks",
     "HookKeyframes",
     "TimestepsRange",
@@ -2291,6 +2343,7 @@ __all__ = [
     "FlowControl",
     "Accumulation",
     "Load3DCamera",
+    "Load3DModelInfo",
     "Load3D",
     "Load3DAnimation",
     "Photomaker",

comfy_api/latest/_ui.py

@@ -452,6 +452,16 @@ class PreviewUI3D(_UIOutput):
         return {"result": [self.model_file, self.camera_info, self.bg_image_path]}
 
 
+class PreviewUI3DAdvanced(_UIOutput):
+    def __init__(self, model_file, camera_info, model_3d_info):
+        self.model_file = model_file
+        self.camera_info = camera_info
+        self.model_3d_info = model_3d_info
+
+    def as_dict(self):
+        return {"result": [self.model_file, self.camera_info, self.model_3d_info]}
+
+
 class PreviewText(_UIOutput):
     def __init__(self, value: str, **kwargs):
         self.value = value
@@ -471,5 +481,6 @@ __all__ = [
     "PreviewAudio",
     "PreviewVideo",
     "PreviewUI3D",
+    "PreviewUI3DAdvanced",
     "PreviewText",
 ]

comfy_api/latest/_util/__init__.py

@@ -1,5 +1,5 @@
 from .video_types import VideoContainer, VideoCodec, VideoComponents
-from .geometry_types import VOXEL, MESH, File3D
+from .geometry_types import VOXEL, MESH, SPLAT, File3D
 from .image_types import SVG
 
 __all__ = [
@@ -9,6 +9,7 @@ __all__ = [
     "VideoComponents",
     "VOXEL",
     "MESH",
+    "SPLAT",
     "File3D",
     "SVG",
 ]

comfy_api/latest/_util/geometry_types.py

@@ -11,13 +11,32 @@ class VOXEL:
         self.data = data
 
 
+class SPLAT:
+    """A batch of 3D Gaussian splats in render-ready (activated, world-space) form.
+
+    Tensors are (B, N, ...) and zero-padded to a common N across the batch; `counts` (B,) holds the
+    real per-item lengths (None when rows are uniform and no slicing is needed). SH coefficients are
+    stored as (B, N, K, 3) with K = (sh_degree + 1)**2; the DC (diffuse) term is sh[..., 0, :].
+    """
+
+    def __init__(self, positions: torch.Tensor, scales: torch.Tensor, rotations: torch.Tensor,
+                 opacities: torch.Tensor, sh: torch.Tensor, counts: torch.Tensor | None = None):
+        self.positions = positions    # (B, N, 3) world-space centers
+        self.scales = scales          # (B, N, 3) linear (positive) per-axis std
+        self.rotations = rotations    # (B, N, 4) quaternion wxyz (normalized)
+        self.opacities = opacities    # (B, N, 1) in [0, 1]
+        self.sh = sh                  # (B, N, K, 3) spherical-harmonic color coefficients
+        self.counts = counts          # (B,) real lengths, or None
+
+
 class MESH:
     def __init__(self, vertices: torch.Tensor, faces: torch.Tensor,
                  uvs: torch.Tensor | None = None,
                  vertex_colors: torch.Tensor | None = None,
                  texture: torch.Tensor | None = None,
                  vertex_counts: torch.Tensor | None = None,
-                 face_counts: torch.Tensor | None = None):
+                 face_counts: torch.Tensor | None = None,
+                 unlit: bool = False):
 
         assert (vertex_counts is None) == (face_counts is None), \
             "vertex_counts and face_counts must be provided together (both or neither)"
@@ -30,6 +49,8 @@ class MESH:
         # these hold the real per-item lengths (B,). None means rows are uniform and no slicing is needed.
         self.vertex_counts = vertex_counts
         self.face_counts = face_counts
+        # Render flat / emissive (no scene lighting) when saved, e.g. for gaussian-splat-derived meshes.
+        self.unlit = unlit
 
 
 class File3D:

comfy_api_nodes/apis/beeble.py

@@ -0,0 +1,32 @@
+from pydantic import BaseModel, Field
+
+
+class CreateSwitchXRequest(BaseModel):
+    generation_type: str = Field(...)
+    source_uri: str = Field(...)
+    alpha_mode: str = Field(...)
+    prompt: str | None = Field(None, max_length=2000)
+    reference_image_uri: str | None = Field(None)
+    alpha_uri: str | None = Field(None)
+    max_resolution: int = Field(1080)
+    callback_url: str | None = Field(None)
+    idempotency_key: str | None = Field(None, max_length=256, min_length=1)
+
+
+class SwitchXOutputUrls(BaseModel):
+    render: str | None = Field(None)
+    source: str | None = Field(None)
+    alpha: str | None = Field(None)
+
+
+class SwitchXStatusResponse(BaseModel):
+    id: str = Field(...)
+    status: str = Field(...)
+    progress: int | None = Field(None)
+    generation_type: str | None = Field(None)
+    alpha_mode: str | None = Field(None)
+    output: SwitchXOutputUrls | None = Field(None)
+    error: str | None = Field(None)
+    created_at: str | None = Field(None)
+    modified_at: str | None = Field(None)
+    completed_at: str | None = Field(None)

comfy_api_nodes/apis/bytedance.py

@@ -158,8 +158,9 @@ class SeedanceCreateAssetResponse(BaseModel):
 
 
 class SeedanceVirtualLibraryCreateAssetRequest(BaseModel):
-    url: str = Field(..., description="Publicly accessible URL of the image asset to upload.")
+    url: str = Field(..., description="Publicly accessible URL of the asset to upload.")
     hash: str = Field(..., description="Dedup key. Re-submitting the same hash returns the existing asset id.")
+    asset_type: str | None = Field(None, description="BytePlus asset type. Defaults to Image server-side when omitted.")
 
 
 # Dollars per 1K tokens, keyed by (model_id, has_video_input).

comfy_api_nodes/apis/krea.py

@@ -0,0 +1,46 @@
+"""Pydantic models for the Krea image-generation API."""
+
+from pydantic import BaseModel, Field
+
+
+class KreaMoodboard(BaseModel):
+    id: str = Field(...)
+    strength: float = Field(default=0.35, ge=-0.5, le=1.5)
+
+
+class KreaImageStyleReference(BaseModel):
+    strength: float = Field(..., ge=-2.0, le=2.0)
+    url: str | None = Field(default=None)
+
+
+class KreaGenerateImageRequest(BaseModel):
+    prompt: str = Field(...)
+    aspect_ratio: str = Field(...)
+    resolution: str = Field(...)
+    seed: int | None = Field(default=None)
+    creativity: str = Field(default="medium")
+    moodboards: list[KreaMoodboard] | None = Field(default=None)
+    image_style_references: list[KreaImageStyleReference] | None = Field(default=None)
+
+
+class KreaJobResult(BaseModel):
+    urls: list[str] | None = Field(default=None)
+    style_id: str | None = Field(default=None)
+
+
+class KreaJob(BaseModel):
+    job_id: str = Field(...)
+    status: str = Field(...)
+    created_at: str = Field(...)
+    completed_at: str | None = Field(default=None)
+    result: KreaJobResult | None = Field(default=None)
+
+
+class KreaAssetResponse(BaseModel):
+    id: str = Field(...)
+    image_url: str = Field(...)
+    uploaded_at: str = Field(...)
+    width: float | None = Field(default=None)
+    height: float | None = Field(default=None)
+    size_bytes: float | None = Field(default=None)
+    mime_type: str | None = Field(default=None)

comfy_api_nodes/apis/tripo.py

@@ -1,25 +1,25 @@
 from enum import Enum
-from typing import Optional, Any
+from typing import Any
 
 from pydantic import BaseModel, Field, RootModel
 
 
 class TripoModelVersion(str, Enum):
-    v3_1_20260211 = 'v3.1-20260211'
-    v3_0_20250812 = 'v3.0-20250812'
-    v2_5_20250123 = 'v2.5-20250123'
-    v2_0_20240919 = 'v2.0-20240919'
-    v1_4_20240625 = 'v1.4-20240625'
+    v3_1_20260211 = "v3.1-20260211"
+    v3_0_20250812 = "v3.0-20250812"
+    v2_5_20250123 = "v2.5-20250123"
+    v2_0_20240919 = "v2.0-20240919"
+    v1_4_20240625 = "v1.4-20240625"
 
 
 class TripoGeometryQuality(str, Enum):
-    standard = 'standard'
-    detailed = 'detailed'
+    standard = "standard"
+    detailed = "detailed"
 
 
 class TripoTextureQuality(str, Enum):
-    standard = 'standard'
-    detailed = 'detailed'
+    standard = "standard"
+    detailed = "detailed"
 
 
 class TripoStyle(str, Enum):
@@ -33,6 +33,7 @@ class TripoStyle(str, Enum):
     ANCIENT_BRONZE = "ancient_bronze"
     NONE = "None"
 
+
 class TripoTaskType(str, Enum):
     TEXT_TO_MODEL = "text_to_model"
     IMAGE_TO_MODEL = "image_to_model"
@@ -45,26 +46,27 @@ class TripoTaskType(str, Enum):
     STYLIZE_MODEL = "stylize_model"
     CONVERT_MODEL = "convert_model"
 
+
 class TripoTextureAlignment(str, Enum):
     ORIGINAL_IMAGE = "original_image"
     GEOMETRY = "geometry"
 
+
 class TripoOrientation(str, Enum):
     ALIGN_IMAGE = "align_image"
     DEFAULT = "default"
 
+
 class TripoOutFormat(str, Enum):
     GLB = "glb"
     FBX = "fbx"
 
-class TripoTopology(str, Enum):
-    BIP = "bip"
-    QUAD = "quad"
 
 class TripoSpec(str, Enum):
     MIXAMO = "mixamo"
     TRIPO = "tripo"
 
+
 class TripoAnimation(str, Enum):
     IDLE = "preset:idle"
     WALK = "preset:walk"
@@ -83,11 +85,6 @@ class TripoAnimation(str, Enum):
     SERPENTINE_MARCH = "preset:serpentine:march"
     AQUATIC_MARCH = "preset:aquatic:march"
 
-class TripoStylizeStyle(str, Enum):
-    LEGO = "lego"
-    VOXEL = "voxel"
-    VORONOI = "voronoi"
-    MINECRAFT = "minecraft"
 
 class TripoConvertFormat(str, Enum):
     GLTF = "GLTF"
@@ -97,6 +94,7 @@ class TripoConvertFormat(str, Enum):
     STL = "STL"
     _3MF = "3MF"
 
+
 class TripoTextureFormat(str, Enum):
     BMP = "BMP"
     DPX = "DPX"
@@ -108,6 +106,7 @@ class TripoTextureFormat(str, Enum):
     TIFF = "TIFF"
     WEBP = "WEBP"
 
+
 class TripoTaskStatus(str, Enum):
     QUEUED = "queued"
     RUNNING = "running"
@@ -118,183 +117,223 @@ class TripoTaskStatus(str, Enum):
     BANNED = "banned"
     EXPIRED = "expired"
 
+
 class TripoFbxPreset(str, Enum):
     BLENDER = "blender"
     MIXAMO = "mixamo"
     _3DSMAX = "3dsmax"
 
+
 class TripoFileTokenReference(BaseModel):
-    type: Optional[str] = Field(None, description='The type of the reference')
+    type: str | None = Field(None, description="The type of the reference")
     file_token: str
 
+
 class TripoUrlReference(BaseModel):
-    type: Optional[str] = Field(None, description='The type of the reference')
+    type: str | None = Field(None, description="The type of the reference")
     url: str
 
+
 class TripoObjectStorage(BaseModel):
     bucket: str
     key: str
 
+
 class TripoObjectReference(BaseModel):
     type: str
     object: TripoObjectStorage
 
+
 class TripoFileEmptyReference(BaseModel):
     pass
 
+
 class TripoFileReference(RootModel):
     root: TripoFileTokenReference | TripoUrlReference | TripoObjectReference | TripoFileEmptyReference
 
-class TripoGetStsTokenRequest(BaseModel):
-    format: str = Field(..., description='The format of the image')
 
 class TripoTextToModelRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.TEXT_TO_MODEL, description='Type of task')
-    prompt: str = Field(..., description='The text prompt describing the model to generate', max_length=1024)
-    negative_prompt: Optional[str] = Field(None, description='The negative text prompt', max_length=1024)
-    model_version: Optional[TripoModelVersion] = TripoModelVersion.v2_5_20250123
-    face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to')
-    texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model')
-    pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model')
-    image_seed: Optional[int] = Field(None, description='The seed for the text')
-    model_seed: Optional[int] = Field(None, description='The seed for the model')
-    texture_seed: Optional[int] = Field(None, description='The seed for the texture')
-    texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard
-    geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard
-    style: Optional[TripoStyle] = None
-    auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model')
-    quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model')
+    type: TripoTaskType = Field(TripoTaskType.TEXT_TO_MODEL, description="Type of task")
+    prompt: str = Field(..., description="The text prompt describing the model to generate", max_length=1024)
+    negative_prompt: str | None = Field(None, description="The negative text prompt", max_length=1024)
+    model_version: TripoModelVersion | None = TripoModelVersion.v2_5_20250123
+    face_limit: int | None = Field(None, description="The number of faces to limit the generation to")
+    texture: bool | None = Field(True, description="Whether to apply texture to the generated model")
+    pbr: bool | None = Field(True, description="Whether to apply PBR to the generated model")
+    image_seed: int | None = Field(None, description="The seed for the text")
+    model_seed: int | None = Field(None, description="The seed for the model")
+    texture_seed: int | None = Field(None, description="The seed for the texture")
+    texture_quality: TripoTextureQuality | None = TripoTextureQuality.standard
+    geometry_quality: TripoGeometryQuality | None = TripoGeometryQuality.standard
+    style: TripoStyle | None = None
+    auto_size: bool | None = Field(False, description="Whether to auto-size the model")
+    quad: bool | None = Field(False, description="Whether to apply quad to the generated model")
+
 
 class TripoImageToModelRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.IMAGE_TO_MODEL, description='Type of task')
-    file: TripoFileReference = Field(..., description='The file reference to convert to a model')
-    model_version: Optional[TripoModelVersion] = Field(None, description='The model version to use for generation')
-    face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to')
-    texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model')
-    pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model')
-    model_seed: Optional[int] = Field(None, description='The seed for the model')
-    texture_seed: Optional[int] = Field(None, description='The seed for the texture')
-    texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard
-    geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard
-    texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method')
-    style: Optional[TripoStyle] = Field(None, description='The style to apply to the generated model')
-    auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model')
-    orientation: Optional[TripoOrientation] = TripoOrientation.DEFAULT
-    quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model')
+    type: TripoTaskType = Field(TripoTaskType.IMAGE_TO_MODEL, description="Type of task")
+    file: TripoFileReference = Field(..., description="The file reference to convert to a model")
+    model_version: TripoModelVersion | None = Field(None, description="The model version to use for generation")
+    face_limit: int | None = Field(None, description="The number of faces to limit the generation to")
+    texture: bool | None = Field(True, description="Whether to apply texture to the generated model")
+    pbr: bool | None = Field(True, description="Whether to apply PBR to the generated model")
+    model_seed: int | None = Field(None, description="The seed for the model")
+    texture_seed: int | None = Field(None, description="The seed for the texture")
+    texture_quality: TripoTextureQuality | None = TripoTextureQuality.standard
+    geometry_quality: TripoGeometryQuality | None = TripoGeometryQuality.standard
+    texture_alignment: TripoTextureAlignment | None = Field(
+        TripoTextureAlignment.ORIGINAL_IMAGE, description="The texture alignment method"
+    )
+    style: TripoStyle | None = Field(None, description="The style to apply to the generated model")
+    auto_size: bool | None = Field(False, description="Whether to auto-size the model")
+    orientation: TripoOrientation | None = TripoOrientation.DEFAULT
+    quad: bool | None = Field(False, description="Whether to apply quad to the generated model")
+
 
 class TripoMultiviewToModelRequest(BaseModel):
     type: TripoTaskType = TripoTaskType.MULTIVIEW_TO_MODEL
-    files: list[TripoFileReference] = Field(..., description='The file references to convert to a model')
-    model_version: Optional[TripoModelVersion] = Field(None, description='The model version to use for generation')
-    orthographic_projection: Optional[bool] = Field(False, description='Whether to use orthographic projection')
-    face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to')
-    texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model')
-    pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model')
-    model_seed: Optional[int] = Field(None, description='The seed for the model')
-    texture_seed: Optional[int] = Field(None, description='The seed for the texture')
-    texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard
-    geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard
-    texture_alignment: Optional[TripoTextureAlignment] = TripoTextureAlignment.ORIGINAL_IMAGE
-    auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model')
-    orientation: Optional[TripoOrientation] = Field(TripoOrientation.DEFAULT, description='The orientation for the model')
-    quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model')
+    files: list[TripoFileReference] = Field(..., description="The file references to convert to a model")
+    model_version: TripoModelVersion | None = Field(None, description="The model version to use for generation")
+    orthographic_projection: bool | None = Field(False, description="Whether to use orthographic projection")
+    face_limit: int | None = Field(None, description="The number of faces to limit the generation to")
+    texture: bool | None = Field(True, description="Whether to apply texture to the generated model")
+    pbr: bool | None = Field(True, description="Whether to apply PBR to the generated model")
+    model_seed: int | None = Field(None, description="The seed for the model")
+    texture_seed: int | None = Field(None, description="The seed for the texture")
+    texture_quality: TripoTextureQuality | None = TripoTextureQuality.standard
+    geometry_quality: TripoGeometryQuality | None = TripoGeometryQuality.standard
+    texture_alignment: TripoTextureAlignment | None = TripoTextureAlignment.ORIGINAL_IMAGE
+    auto_size: bool | None = Field(False, description="Whether to auto-size the model")
+    orientation: TripoOrientation | None = Field(TripoOrientation.DEFAULT, description="The orientation for the model")
+    quad: bool | None = Field(False, description="Whether to apply quad to the generated model")
+
 
 class TripoTextureModelRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.TEXTURE_MODEL, description='Type of task')
-    original_model_task_id: str = Field(..., description='The task ID of the original model')
-    texture: Optional[bool] = Field(True, description='Whether to apply texture to the model')
-    pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the model')
-    model_seed: Optional[int] = Field(None, description='The seed for the model')
-    texture_seed: Optional[int] = Field(None, description='The seed for the texture')
-    texture_quality: Optional[TripoTextureQuality] = Field(None, description='The quality of the texture')
-    texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method')
+    type: TripoTaskType = Field(TripoTaskType.TEXTURE_MODEL, description="Type of task")
+    original_model_task_id: str = Field(..., description="The task ID of the original model")
+    texture: bool | None = Field(True, description="Whether to apply texture to the model")
+    pbr: bool | None = Field(True, description="Whether to apply PBR to the model")
+    model_seed: int | None = Field(None, description="The seed for the model")
+    texture_seed: int | None = Field(None, description="The seed for the texture")
+    texture_quality: TripoTextureQuality | None = Field(None, description="The quality of the texture")
+    texture_alignment: TripoTextureAlignment | None = Field(
+        TripoTextureAlignment.ORIGINAL_IMAGE, description="The texture alignment method"
+    )
+
 
 class TripoRefineModelRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.REFINE_MODEL, description='Type of task')
-    draft_model_task_id: str = Field(..., description='The task ID of the draft model')
+    type: TripoTaskType = Field(TripoTaskType.REFINE_MODEL, description="Type of task")
+    draft_model_task_id: str = Field(..., description="The task ID of the draft model")
 
-class TripoAnimatePrerigcheckRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.ANIMATE_PRERIGCHECK, description='Type of task')
-    original_model_task_id: str = Field(..., description='The task ID of the original model')
 
 class TripoAnimateRigRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.ANIMATE_RIG, description='Type of task')
-    original_model_task_id: str = Field(..., description='The task ID of the original model')
-    out_format: Optional[TripoOutFormat] = Field(TripoOutFormat.GLB, description='The output format')
-    spec: Optional[TripoSpec] = Field(TripoSpec.TRIPO, description='The specification for rigging')
+    type: TripoTaskType = Field(TripoTaskType.ANIMATE_RIG, description="Type of task")
+    original_model_task_id: str = Field(..., description="The task ID of the original model")
+    out_format: TripoOutFormat | None = Field(TripoOutFormat.GLB, description="The output format")
+    spec: TripoSpec | None = Field(TripoSpec.TRIPO, description="The specification for rigging")
+
 
 class TripoAnimateRetargetRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.ANIMATE_RETARGET, description='Type of task')
-    original_model_task_id: str = Field(..., description='The task ID of the original model')
-    animation: TripoAnimation = Field(..., description='The animation to apply')
-    out_format: Optional[TripoOutFormat] = Field(TripoOutFormat.GLB, description='The output format')
-    bake_animation: Optional[bool] = Field(True, description='Whether to bake the animation')
+    type: TripoTaskType = Field(TripoTaskType.ANIMATE_RETARGET, description="Type of task")
+    original_model_task_id: str = Field(..., description="The task ID of the original model")
+    animation: TripoAnimation = Field(..., description="The animation to apply")
+    out_format: TripoOutFormat | None = Field(TripoOutFormat.GLB, description="The output format")
+    bake_animation: bool | None = Field(True, description="Whether to bake the animation")
 
-class TripoStylizeModelRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.STYLIZE_MODEL, description='Type of task')
-    style: TripoStylizeStyle = Field(..., description='The style to apply to the model')
-    original_model_task_id: str = Field(..., description='The task ID of the original model')
-    block_size: Optional[int] = Field(80, description='The block size for stylization')
 
 class TripoConvertModelRequest(BaseModel):
-    type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description='Type of task')
-    format: TripoConvertFormat = Field(..., description='The format to convert to')
-    original_model_task_id: str = Field(..., description='The task ID of the original model')
-    quad: Optional[bool] = Field(None, description='Whether to apply quad to the model')
-    force_symmetry: Optional[bool] = Field(None, description='Whether to force symmetry')
-    face_limit: Optional[int] = Field(None, description='The number of faces to limit the conversion to')
-    flatten_bottom: Optional[bool] = Field(None, description='Whether to flatten the bottom of the model')
-    flatten_bottom_threshold: Optional[float] = Field(None, description='The threshold for flattening the bottom')
-    texture_size: Optional[int] = Field(None, description='The size of the texture')
-    texture_format: Optional[TripoTextureFormat] = Field(TripoTextureFormat.JPEG, description='The format of the texture')
-    pivot_to_center_bottom: Optional[bool] = Field(None, description='Whether to pivot to the center bottom')
-    scale_factor: Optional[float] = Field(None, description='The scale factor for the model')
-    with_animation: Optional[bool] = Field(None, description='Whether to include animations')
-    pack_uv: Optional[bool] = Field(None, description='Whether to pack the UVs')
-    bake: Optional[bool] = Field(None, description='Whether to bake the model')
-    part_names: Optional[list[str]] = Field(None, description='The names of the parts to include')
-    fbx_preset: Optional[TripoFbxPreset] = Field(None, description='The preset for the FBX export')
-    export_vertex_colors: Optional[bool] = Field(None, description='Whether to export the vertex colors')
-    export_orientation: Optional[TripoOrientation] = Field(None, description='The orientation for the export')
-    animate_in_place: Optional[bool] = Field(None, description='Whether to animate in place')
+    type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description="Type of task")
+    format: TripoConvertFormat = Field(..., description="The format to convert to")
+    original_model_task_id: str = Field(..., description="The task ID of the original model")
+    quad: bool | None = Field(None, description="Whether to apply quad to the model")
+    force_symmetry: bool | None = Field(None, description="Whether to force symmetry")
+    face_limit: int | None = Field(None, description="The number of faces to limit the conversion to")
+    flatten_bottom: bool | None = Field(None, description="Whether to flatten the bottom of the model")
+    flatten_bottom_threshold: float | None = Field(None, description="The threshold for flattening the bottom")
+    texture_size: int | None = Field(None, description="The size of the texture")
+    texture_format: TripoTextureFormat | None = Field(TripoTextureFormat.JPEG, description="The format of the texture")
+    pivot_to_center_bottom: bool | None = Field(None, description="Whether to pivot to the center bottom")
+    scale_factor: float | None = Field(None, description="The scale factor for the model")
+    with_animation: bool | None = Field(None, description="Whether to include animations")
+    pack_uv: bool | None = Field(None, description="Whether to pack the UVs")
+    bake: bool | None = Field(None, description="Whether to bake the model")
+    part_names: list[str] | None = Field(None, description="The names of the parts to include")
+    fbx_preset: TripoFbxPreset | None = Field(None, description="The preset for the FBX export")
+    export_vertex_colors: bool | None = Field(None, description="Whether to export the vertex colors")
+    export_orientation: TripoOrientation | None = Field(None, description="The orientation for the export")
+    animate_in_place: bool | None = Field(None, description="Whether to animate in place")
+
+
+class TripoP1CommonRequest(BaseModel):
+    """Fields supported by Tripo P1 across all input types."""
+
+    model_version: str = Field("P1-20260311")
+    model_seed: int | None = Field(None, description="Random seed for geometry generation")
+    face_limit: int | None = Field(None, ge=48, le=20000, description="Target face count (48-20000)")
+    texture: bool | None = Field(None, description="Enable texturing; pbr=True forces this true")
+    pbr: bool | None = Field(None, description="Enable PBR maps; when true, texture is also enabled")
+    texture_seed: int | None = Field(None, description="Random seed for texture generation")
+    texture_quality: str | None = Field(None, description='"standard" or "detailed"')
+    auto_size: bool | None = Field(None, description="Scale to real-world meters")
+    compress: str | None = Field(None, description='Only "geometry" is supported')
+    export_uv: bool | None = Field(None, description="Perform UV unwrapping during generation")
+
+
+class TripoP1TextToModelRequest(TripoP1CommonRequest):
+    type: str = "text_to_model"
+    prompt: str = Field(..., max_length=1024)
+    negative_prompt: str | None = Field(None, max_length=255)
+    image_seed: int | None = None
+
+
+class TripoP1ImageToModelRequest(TripoP1CommonRequest):
+    type: str = "image_to_model"
+    file: TripoFileReference
+    enable_image_autofix: bool | None = None
+    texture_alignment: str | None = Field(None, description='"original_image" or "geometry"')
+    orientation: str | None = Field(None, description='"default" or "align_image"; needs texture=true')
+
+
+class TripoP1MultiviewToModelRequest(TripoP1CommonRequest):
+    """P1 multiview generation.
+
+    Tripo requires `files` to be exactly four entries in [front, left, back, right] order with `{}`
+    (TripoFileEmptyReference) for omitted slots; front is required and at least two images total must be provided.
+    """
+
+    type: str = "multiview_to_model"
+    files: list[TripoFileReference]
+    texture_alignment: str | None = None
+    orientation: str | None = None
 
 
 class TripoTaskOutput(BaseModel):
-    model: Optional[str] = Field(None, description='URL to the model')
-    base_model: Optional[str] = Field(None, description='URL to the base model')
-    pbr_model: Optional[str] = Field(None, description='URL to the PBR model')
-    rendered_image: Optional[str] = Field(None, description='URL to the rendered image')
-    riggable: Optional[bool] = Field(None, description='Whether the model is riggable')
+    model: str | None = Field(None, description="URL to the model")
+    base_model: str | None = Field(None, description="URL to the base model")
+    pbr_model: str | None = Field(None, description="URL to the PBR model")
+    rendered_image: str | None = Field(None, description="URL to the rendered image")
+    riggable: bool | None = Field(None, description="Whether the model is riggable")
+
 
 class TripoTask(BaseModel):
-    task_id: str = Field(..., description='The task ID')
-    type: Optional[str] = Field(None, description='The type of task')
-    status: Optional[TripoTaskStatus] = Field(None, description='The status of the task')
-    input: Optional[dict[str, Any]] = Field(None, description='The input parameters for the task')
-    output: Optional[TripoTaskOutput] = Field(None, description='The output of the task')
-    progress: Optional[int] = Field(None, description='The progress of the task', ge=0, le=100)
-    create_time: Optional[int] = Field(None, description='The creation time of the task')
-    running_left_time: Optional[int] = Field(None, description='The estimated time left for the task')
-    queue_position: Optional[int] = Field(None, description='The position in the queue')
+    task_id: str = Field(..., description="The task ID")
+    type: str | None = Field(None, description="The type of task")
+    status: TripoTaskStatus | None = Field(None, description="The status of the task")
+    input: dict[str, Any] | None = Field(None, description="The input parameters for the task")
+    output: TripoTaskOutput | None = Field(None, description="The output of the task")
+    progress: int | None = Field(None, description="The progress of the task", ge=0, le=100)
+    create_time: int | None = Field(None, description="The creation time of the task")
+    running_left_time: int | None = Field(None, description="The estimated time left for the task")
+    queue_position: int | None = Field(None, description="The position in the queue")
     consumed_credit: int | None = Field(None)
 
+
 class TripoTaskResponse(BaseModel):
-    code: int = Field(0, description='The response code')
-    data: TripoTask = Field(..., description='The task data')
+    code: int = Field(0, description="The response code")
+    data: TripoTask = Field(..., description="The task data")
 
-class TripoGeneralResponse(BaseModel):
-    code: int = Field(0, description='The response code')
-    data: dict[str, str] = Field(..., description='The task ID data')
-
-class TripoBalanceData(BaseModel):
-    balance: float = Field(..., description='The account balance')
-    frozen: float = Field(..., description='The frozen balance')
-
-class TripoBalanceResponse(BaseModel):
-    code: int = Field(0, description='The response code')
-    data: TripoBalanceData = Field(..., description='The balance data')
 
 class TripoErrorResponse(BaseModel):
-    code: int = Field(..., description='The error code')
-    message: str = Field(..., description='The error message')
-    suggestion: str = Field(..., description='The suggestion for fixing the error')
+    code: int = Field(..., description="The error code")
+    message: str = Field(..., description="The error message")
+    suggestion: str = Field(..., description="The suggestion for fixing the error")

comfy_api_nodes/nodes_anthropic.py

@@ -155,7 +155,7 @@ class ClaudeNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ClaudeNode",
             display_name="Anthropic Claude",
-            category="api node/text/Anthropic",
+            category="text/partner/Anthropic",
             essentials_category="Text Generation",
             description="Generate text responses with Anthropic's Claude models. "
             "Provide a text prompt and optionally one or more images for multimodal context.",

comfy_api_nodes/nodes_beeble.py

@@ -0,0 +1,404 @@
+from fractions import Fraction
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input, InputImpl, Types
+from comfy_api_nodes.apis.beeble import (
+    CreateSwitchXRequest,
+    SwitchXStatusResponse,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    bytesio_to_image_tensor,
+    convert_mask_to_image,
+    download_url_as_bytesio,
+    download_url_to_image_tensor,
+    download_url_to_video_output,
+    downscale_image_tensor,
+    downscale_video_to_max_pixels,
+    poll_op,
+    sync_op,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_string,
+    validate_video_frame_count,
+)
+
+_MAX_PIXELS = 2_770_000
+_MAX_FRAMES = 240
+_MAX_PROMPT_LEN = 2000
+
+
+def _validate_inputs(prompt: str | None, reference_image: Input.Image | None) -> str | None:
+    """Beeble requires at least one of prompt or reference_image. Returns the cleaned prompt."""
+    cleaned = prompt.strip() if prompt else ""
+    if not cleaned and reference_image is None:
+        raise ValueError("At least one of 'prompt' or 'reference_image' must be provided.")
+    if cleaned:
+        validate_string(cleaned, strip_whitespace=False, max_length=_MAX_PROMPT_LEN)
+    return cleaned or None
+
+
+async def _upload_mask_as_image(
+    cls: type[IO.ComfyNode],
+    mask: Input.Image,
+    *,
+    wait_label: str,
+) -> str:
+    """Encode a single-frame MASK (H, W) or (1, H, W) as a PNG and upload."""
+    if mask.dim() == 2:
+        mask = mask.unsqueeze(0)
+    image = convert_mask_to_image(mask[:1])
+    return await upload_image_to_comfyapi(
+        cls,
+        image,
+        mime_type="image/png",
+        wait_label=wait_label,
+        total_pixels=_MAX_PIXELS,
+    )
+
+
+async def _upload_mask_batch_as_video(
+    cls: type[IO.ComfyNode],
+    mask: Input.Image,
+    *,
+    frame_rate: Fraction,
+    source_frame_count: int,
+    wait_label: str,
+) -> str:
+    """Encode a MASK batch (N, H, W) as a grayscale H.264 MP4 at frame_rate and upload.
+
+    The matte is always downscaled to the pixel budget so it stays within Beeble's limit and
+    keeps the same dimensions as the (similarly downscaled) source — both use the same algorithm
+    from the same starting dimensions, and downscaling is a no-op when already within budget.
+    """
+    if mask.dim() == 2:
+        mask = mask.unsqueeze(0)
+    if mask.shape[0] != source_frame_count:
+        raise ValueError(
+            f"Custom alpha video frame count ({mask.shape[0]}) does not match the "
+            f"source video frame count ({source_frame_count}). The Beeble API requires "
+            "one mask per source frame."
+        )
+    images = downscale_image_tensor(convert_mask_to_image(mask), _MAX_PIXELS)
+    alpha_video = InputImpl.VideoFromComponents(Types.VideoComponents(images=images, audio=None, frame_rate=frame_rate))
+    return await upload_video_to_comfyapi(cls, alpha_video, wait_label=wait_label)
+
+
+def _alpha_mode_input(*, video: bool) -> IO.DynamicCombo.Input:
+    """Build the alpha_mode DynamicCombo with mode-specific extra inputs."""
+    select_keyframe_tooltip = (
+        "First-frame keyframe mask. Beeble propagates this across the video." if video else "Grayscale keyframe mask."
+    )
+    custom_tooltip = (
+        "Per-frame grayscale mask covering the entire video. "
+        "Must have the same frame count as the source. "
+        "Connect a MASK output from SAM3_TrackToMask or similar."
+        if video
+        else "Grayscale mask to apply."
+    )
+    return IO.DynamicCombo.Input(
+        "alpha_mode",
+        tooltip=(
+            "Controls how SwitchX decides what to keep vs. regenerate. "
+            "'auto' isolates the main subject automatically. "
+            "'fill' regenerates the entire frame while preserving geometry. "
+            "'select' propagates a first-frame keyframe across the clip. "
+            "'custom' uses a per-frame alpha matte you provide."
+        ),
+        options=[
+            IO.DynamicCombo.Option("auto", []),
+            IO.DynamicCombo.Option("fill", []),
+            IO.DynamicCombo.Option(
+                "select",
+                [IO.Mask.Input("alpha_keyframe", tooltip=select_keyframe_tooltip)],
+            ),
+            IO.DynamicCombo.Option(
+                "custom",
+                [IO.Mask.Input("alpha_mask", tooltip=custom_tooltip)],
+            ),
+        ],
+    )
+
+
+def _common_inputs(*, source: IO.Input, video: bool) -> list[IO.Input]:
+    return [
+        source,
+        IO.String.Input(
+            "prompt",
+            multiline=True,
+            default="",
+            tooltip=(
+                "Text description of the desired output (max 2000 chars). "
+                "At least one of 'prompt' or 'reference_image' is required."
+            ),
+        ),
+        IO.Image.Input(
+            "reference_image",
+            optional=True,
+            tooltip=(
+                "Reference image whose look (background, lighting, costume) the result "
+                "should adopt. At least one of 'reference_image' or 'prompt' is required."
+            ),
+        ),
+        _alpha_mode_input(video=video),
+        IO.Combo.Input(
+            "max_resolution",
+            options=["1080p", "720p"],
+            default="1080p",
+            tooltip="Maximum output resolution.",
+        ),
+        IO.Int.Input(
+            "seed",
+            default=0,
+            min=0,
+            max=2147483647,
+            control_after_generate=True,
+            tooltip=(
+                "Seed controls whether the node should re-run; " "results are non-deterministic regardless of seed."
+            ),
+        ),
+    ]
+
+
+async def _submit_and_poll(
+    cls: type[IO.ComfyNode],
+    request: CreateSwitchXRequest,
+) -> SwitchXStatusResponse:
+    initial = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/beeble/v1/switchx/generations", method="POST"),
+        response_model=SwitchXStatusResponse,
+        data=request,
+    )
+    return await poll_op(
+        cls,
+        ApiEndpoint(path=f"/proxy/beeble/v1/switchx/generations/{initial.id}"),
+        response_model=SwitchXStatusResponse,
+        status_extractor=lambda r: r.status,
+        progress_extractor=lambda r: r.progress,
+    )
+
+
+def _require_output_url(response: SwitchXStatusResponse, name: str) -> str:
+    if response.output is None or getattr(response.output, name) is None:
+        raise RuntimeError(f"Beeble job {response.id} completed without a {name!r} output URL.")
+    return getattr(response.output, name)
+
+
+def _alpha_url(response: SwitchXStatusResponse, mode: str) -> str | None:
+    """URL of the alpha matte, or None when the mode produces no separate matte.
+
+    'fill' selects the whole frame, so Beeble writes no alpha asset even though the status
+    response still returns a (dangling) signed URL for it — fetching it 403s with S3
+    AccessDenied. The other three modes ('auto', 'custom', 'select') all produce a real,
+    downloadable matte.
+    """
+    if mode == "fill" or response.output is None:
+        return None
+    return response.output.alpha
+
+
+class BeebleSwitchXVideoEdit(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="BeebleSwitchXVideoEdit",
+            display_name="Beeble SwitchX Video Edit",
+            category="video/partner/Beeble",
+            description=(
+                "Edit a video with Beeble SwitchX. Switches anything in the scene (background, "
+                "lighting, costume) while preserving the original subject's pixels and motion. "
+                "Provide a reference image and/or text prompt to describe the new look. "
+                "Max 240 frames, max ~2.77MP per frame."
+            ),
+            inputs=_common_inputs(source=IO.Video.Input("video"), video=True),
+            outputs=[
+                IO.Video.Output(display_name="video"),
+                IO.Video.Output(
+                    display_name="alpha",
+                    tooltip="The alpha matte Beeble used. Empty for 'fill' mode, which has no separate matte.",
+                ),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["max_resolution"]),
+                expr="""
+                (
+                  $rate := widgets.max_resolution = "1080p" ? 0.429 : 0.143;
+                  {"type":"usd","usd": $rate, "format":{"suffix":"/30 frames"}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        prompt: str,
+        alpha_mode: dict,
+        max_resolution: str,
+        seed: int,
+        reference_image: Input.Image | None = None,
+    ) -> IO.NodeOutput:
+        cleaned_prompt = _validate_inputs(prompt, reference_image)
+
+        validate_video_frame_count(video, max_frame_count=_MAX_FRAMES)
+        video = downscale_video_to_max_pixels(video, _MAX_PIXELS)
+
+        mode = alpha_mode["alpha_mode"]
+        alpha_uri: str | None = None
+        if mode == "select":
+            alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_keyframe"], wait_label="Uploading keyframe")
+        elif mode == "custom":
+            alpha_uri = await _upload_mask_batch_as_video(
+                cls,
+                alpha_mode["alpha_mask"],
+                frame_rate=video.get_frame_rate(),
+                source_frame_count=video.get_frame_count(),
+                wait_label="Uploading alpha video",
+            )
+
+        source_uri = await upload_video_to_comfyapi(cls, video, wait_label="Uploading source")
+        reference_uri: str | None = None
+        if reference_image is not None:
+            reference_uri = await upload_image_to_comfyapi(
+                cls,
+                reference_image,
+                mime_type="image/png",
+                wait_label="Uploading reference",
+                total_pixels=_MAX_PIXELS,
+            )
+
+        request = CreateSwitchXRequest(
+            generation_type="video",
+            source_uri=source_uri,
+            alpha_mode=mode,
+            prompt=cleaned_prompt,
+            reference_image_uri=reference_uri,
+            alpha_uri=alpha_uri,
+            max_resolution=1080 if max_resolution == "1080p" else 720,
+        )
+        response = await _submit_and_poll(cls, request)
+
+        render = await download_url_to_video_output(_require_output_url(response, "render"))
+        alpha = None
+        if (alpha_url := _alpha_url(response, mode)) is not None:
+            alpha = await download_url_to_video_output(alpha_url)
+        return IO.NodeOutput(render, alpha)
+
+
+class BeebleSwitchXImageEdit(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="BeebleSwitchXImageEdit",
+            display_name="Beeble SwitchX Image Edit",
+            category="image/partner/Beeble",
+            description=(
+                "Edit a single image with Beeble SwitchX. Switches anything in the scene "
+                "(background, lighting, costume) while preserving the original subject's pixels. "
+                "Provide a reference image and/or text prompt to describe the new look. "
+                "Max ~2.77MP."
+            ),
+            inputs=_common_inputs(source=IO.Image.Input("image"), video=False),
+            outputs=[
+                IO.Image.Output(display_name="image"),
+                IO.Mask.Output(
+                    display_name="alpha",
+                    tooltip="The alpha matte Beeble used. Empty for 'fill' mode, which has no separate matte.",
+                ),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["max_resolution"]),
+                expr="""
+                (
+                  $rate := widgets.max_resolution = "1080p" ? 0.429 : 0.143;
+                  {"type":"usd","usd": $rate}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        prompt: str,
+        alpha_mode: dict,
+        max_resolution: str,
+        seed: int,
+        reference_image: Input.Image | None = None,
+    ) -> IO.NodeOutput:
+        cleaned_prompt = _validate_inputs(prompt, reference_image)
+
+        image = downscale_image_tensor(image, _MAX_PIXELS)
+
+        mode = alpha_mode["alpha_mode"]
+        alpha_uri: str | None = None
+        if mode == "select":
+            alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_keyframe"], wait_label="Uploading keyframe")
+        elif mode == "custom":
+            alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_mask"], wait_label="Uploading alpha")
+
+        source_uri = await upload_image_to_comfyapi(
+            cls,
+            image,
+            mime_type="image/png",
+            wait_label="Uploading source",
+            total_pixels=None,
+        )
+        reference_uri: str | None = None
+        if reference_image is not None:
+            reference_uri = await upload_image_to_comfyapi(
+                cls,
+                reference_image,
+                mime_type="image/png",
+                wait_label="Uploading reference",
+                total_pixels=_MAX_PIXELS,
+            )
+
+        request = CreateSwitchXRequest(
+            generation_type="image",
+            source_uri=source_uri,
+            alpha_mode=mode,
+            prompt=cleaned_prompt,
+            reference_image_uri=reference_uri,
+            alpha_uri=alpha_uri,
+            max_resolution=1080 if max_resolution == "1080p" else 720,
+        )
+        response = await _submit_and_poll(cls, request)
+
+        render = await download_url_to_image_tensor(_require_output_url(response, "render"))
+        alpha_mask = None
+        if (alpha_url := _alpha_url(response, mode)) is not None:
+            alpha_image = bytesio_to_image_tensor(await download_url_as_bytesio(alpha_url), mode="L")
+            alpha_mask = alpha_image.squeeze(-1) if alpha_image.dim() == 4 else alpha_image
+        return IO.NodeOutput(render, alpha_mask)
+
+
+class BeebleExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            BeebleSwitchXVideoEdit,
+            BeebleSwitchXImageEdit,
+        ]
+
+
+async def comfy_entrypoint() -> BeebleExtension:
+    return BeebleExtension()

comfy_api_nodes/nodes_bfl.py

@@ -42,7 +42,7 @@ class FluxProUltraImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="FluxProUltraImageNode",
             display_name="Flux 1.1 [pro] Ultra Image",
-            category="api node/image/BFL",
+            category="image/partner/BFL",
             description="Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution.",
             inputs=[
                 IO.String.Input(
@@ -160,7 +160,7 @@ class FluxKontextProImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id=cls.NODE_ID,
             display_name=cls.DISPLAY_NAME,
-            category="api node/image/BFL",
+            category="image/partner/BFL",
             description="Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio.",
             inputs=[
                 IO.String.Input(
@@ -282,7 +282,7 @@ class FluxProExpandNode(IO.ComfyNode):
         return IO.Schema(
             node_id="FluxProExpandNode",
             display_name="Flux.1 Expand Image",
-            category="api node/image/BFL",
+            category="image/partner/BFL",
             description="Outpaints image based on prompt.",
             inputs=[
                 IO.Image.Input("image"),
@@ -419,7 +419,7 @@ class FluxProFillNode(IO.ComfyNode):
         return IO.Schema(
             node_id="FluxProFillNode",
             display_name="Flux.1 Fill Image",
-            category="api node/image/BFL",
+            category="image/partner/BFL",
             description="Inpaints image based on mask and prompt.",
             inputs=[
                 IO.Image.Input("image"),
@@ -545,7 +545,7 @@ class Flux2ProImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id=cls.NODE_ID,
             display_name=cls.DISPLAY_NAME,
-            category="api node/image/BFL",
+            category="image/partner/BFL",
             description="Generates images synchronously based on prompt and resolution.",
             inputs=[
                 IO.String.Input(
@@ -716,7 +716,7 @@ class Flux2ImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Flux2ImageNode",
             display_name="Flux.2 Image",
-            category="api node/image/BFL",
+            category="image/partner/BFL",
             description="Generate images via Flux.2 [pro] or Flux.2 [max] from a prompt and optional reference images.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_bria.py

@@ -31,7 +31,7 @@ class BriaImageEditNode(IO.ComfyNode):
         return IO.Schema(
             node_id="BriaImageEditNode",
             display_name="Bria FIBO Image Edit",
-            category="api node/image/Bria",
+            category="image/partner/Bria",
             description="Edit images using Bria latest model",
             inputs=[
                 IO.Combo.Input("model", options=["FIBO"]),
@@ -169,7 +169,7 @@ class BriaRemoveImageBackground(IO.ComfyNode):
         return IO.Schema(
             node_id="BriaRemoveImageBackground",
             display_name="Bria Remove Image Background",
-            category="api node/image/Bria",
+            category="image/partner/Bria",
             description="Remove the background from an image using Bria RMBG 2.0.",
             inputs=[
                 IO.Image.Input("image"),
@@ -245,7 +245,7 @@ class BriaRemoveVideoBackground(IO.ComfyNode):
         return IO.Schema(
             node_id="BriaRemoveVideoBackground",
             display_name="Bria Remove Video Background",
-            category="api node/video/Bria",
+            category="video/partner/Bria",
             description="Remove the background from a video using Bria. ",
             inputs=[
                 IO.Video.Input("video"),

comfy_api_nodes/nodes_bytedance.py

@@ -2,11 +2,12 @@ import hashlib
 import logging
 import math
 import re
+from io import BytesIO
 
 import torch
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api.latest import IO, ComfyExtension, Input, Types
 from comfy_api_nodes.apis.bytedance import (
     RECOMMENDED_PRESETS,
     RECOMMENDED_PRESETS_SEEDREAM_4,
@@ -43,6 +44,7 @@ from comfy_api_nodes.util import (
     ApiEndpoint,
     download_url_to_image_tensor,
     download_url_to_video_output,
+    downscale_image_tensor_by_max_side,
     downscale_video_to_max_pixels,
     get_number_of_images,
     image_tensor_pair_to_batch,
@@ -121,6 +123,14 @@ def _validate_ref_video_pixels(video: Input.Video, model_id: str, resolution: st
         )
 
 
+def _prepare_seedance_image(image: Input.Image) -> Input.Image:
+    """Auto-downscale a Seedance image input to the per-side limits, then validate it."""
+    validate_image_aspect_ratio(image, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
+    image = downscale_image_tensor_by_max_side(image, max_side=6000)
+    validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000)
+    return image
+
+
 async def _resolve_reference_assets(
     cls: type[IO.ComfyNode],
     asset_ids: list[str],
@@ -308,6 +318,26 @@ async def _seedance_virtual_library_upload_image_asset(
     return f"asset://{create_resp.asset_id}"
 
 
+async def _seedance_virtual_library_upload_video_asset(
+    cls: type[IO.ComfyNode],
+    video: Input.Video,
+    *,
+    wait_label: str = "Uploading video",
+) -> str:
+    buf = BytesIO()
+    video.save_to(buf, format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264)
+    video_hash = hashlib.sha256(buf.getbuffer()).hexdigest()
+    public_url = await upload_video_to_comfyapi(cls, video, wait_label=wait_label)
+    create_resp = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/seedance/virtual-library/assets", method="POST"),
+        response_model=SeedanceCreateAssetResponse,
+        data=SeedanceVirtualLibraryCreateAssetRequest(url=public_url, hash=video_hash, asset_type="Video"),
+    )
+    await _wait_for_asset_active(cls, create_resp.asset_id, group_id="virtual-library")
+    return f"asset://{create_resp.asset_id}"
+
+
 def _seedance2_price_extractor(model_id: str, has_video_input: bool):
     """Returns a price_extractor closure for Seedance 2.0 poll_op."""
     rate = SEEDANCE2_PRICE_PER_1K_TOKENS.get((model_id, has_video_input))
@@ -338,7 +368,7 @@ class ByteDanceImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceImageNode",
             display_name="ByteDance Image",
-            category="api node/image/ByteDance",
+            category="image/partner/ByteDance",
             description="Generate images using ByteDance models via api based on prompt",
             inputs=[
                 IO.Combo.Input("model", options=["seedream-3-0-t2i-250415"]),
@@ -462,7 +492,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceSeedreamNode",
             display_name="ByteDance Seedream 4.5 & 5.0",
-            category="api node/image/ByteDance",
+            category="image/partner/ByteDance",
             description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.",
             inputs=[
                 IO.Combo.Input(
@@ -724,7 +754,7 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceSeedreamNodeV2",
             display_name="ByteDance Seedream 4.5 & 5.0",
-            category="api node/image/ByteDance",
+            category="image/partner/ByteDance",
             description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.",
             inputs=[
                 IO.String.Input(
@@ -890,7 +920,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceTextToVideoNode",
             display_name="ByteDance Text to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate video using ByteDance models via api based on prompt",
             inputs=[
                 IO.Combo.Input(
@@ -1018,7 +1048,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceImageToVideoNode",
             display_name="ByteDance Image to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate video using ByteDance models via api based on image and prompt",
             inputs=[
                 IO.Combo.Input(
@@ -1155,7 +1185,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceFirstLastFrameNode",
             display_name="ByteDance First-Last-Frame to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate video using prompt and first and last frames.",
             inputs=[
                 IO.Combo.Input(
@@ -1303,7 +1333,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceImageReferenceNode",
             display_name="ByteDance Reference Images to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate video using prompt and reference images.",
             inputs=[
                 IO.Combo.Input(
@@ -1546,7 +1576,7 @@ class ByteDance2TextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDance2TextToVideoNode",
             display_name="ByteDance Seedance 2.0 Text to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate video using Seedance 2.0 models based on a text prompt.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1647,7 +1677,7 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDance2FirstLastFrameNode",
             display_name="ByteDance Seedance 2.0 First-Last-Frame to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate video using Seedance 2.0 from a first frame image and optional last frame image.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1760,6 +1790,11 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
         if last_frame is not None and last_frame_asset_id:
             raise ValueError("Provide only one of last_frame or last_frame_asset_id, not both.")
 
+        if first_frame is not None:
+            first_frame = _prepare_seedance_image(first_frame)
+        if last_frame is not None:
+            last_frame = _prepare_seedance_image(last_frame)
+
         asset_ids_to_resolve = [a for a in (first_frame_asset_id, last_frame_asset_id) if a]
         image_assets: dict[str, str] = {}
         if asset_ids_to_resolve:
@@ -1866,7 +1901,7 @@ def _seedance2_reference_inputs(resolutions: list[str], default_ratio: str = "16
         ),
         IO.Boolean.Input(
             "auto_downscale",
-            default=False,
+            default=True,
             optional=True,
             tooltip="Automatically downscale reference videos that exceed the model's pixel budget "
             "for the selected resolution. Aspect ratio is preserved; videos already within limits are untouched.",
@@ -1909,7 +1944,7 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDance2ReferenceNode",
             display_name="ByteDance Seedance 2.0 Reference to Video",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description="Generate, edit, or extend video using Seedance 2.0 with reference images, "
             "videos, and audio. Supports multimodal reference, video editing, and video extension.",
             inputs=[
@@ -2034,6 +2069,9 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
                 f"(audios={len(reference_audios)}, audio assets={len(reference_audio_assets)}). Maximum is 3."
             )
 
+        for key in reference_images:
+            reference_images[key] = _prepare_seedance_image(reference_images[key])
+
         model_id = SEEDANCE_MODELS[model["model"]]
         has_video_input = total_videos > 0
 
@@ -2106,7 +2144,7 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
             content.append(
                 TaskVideoContent(
                     video_url=TaskVideoContentUrl(
-                        url=await upload_video_to_comfyapi(
+                        url=await _seedance_virtual_library_upload_video_asset(
                             cls,
                             reference_videos[key],
                             wait_label=f"Uploading video {i}",
@@ -2203,7 +2241,7 @@ class ByteDanceCreateImageAsset(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceCreateImageAsset",
             display_name="ByteDance Create Image Asset",
-            category="api node/image/ByteDance",
+            category="image/partner/ByteDance",
             description=(
                 "Create a Seedance 2.0 personal image asset. Uploads the input image and "
                 "registers it in the given asset group. If group_id is empty, runs a real-person "
@@ -2270,7 +2308,7 @@ class ByteDanceCreateVideoAsset(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceCreateVideoAsset",
             display_name="ByteDance Create Video Asset",
-            category="api node/video/ByteDance",
+            category="video/partner/ByteDance",
             description=(
                 "Create a Seedance 2.0 personal video asset. Uploads the input video and "
                 "registers it in the given asset group. If group_id is empty, runs a real-person "

comfy_api_nodes/nodes_bytedance_llm.py

@@ -144,7 +144,7 @@ class ByteDanceSeedNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ByteDanceSeedNode",
             display_name="ByteDance Seed",
-            category="api node/text/ByteDance",
+            category="text/partner/ByteDance",
             essentials_category="Text Generation",
             description="Generate text responses with ByteDance's Seed 2.0 models. "
             "Provide a text prompt and optionally one or more images or videos for multimodal context.",

comfy_api_nodes/nodes_elevenlabs.py

@@ -69,7 +69,7 @@ class ElevenLabsSpeechToText(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsSpeechToText",
             display_name="ElevenLabs Speech to Text",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Transcribe audio to text. "
             "Supports automatic language detection, speaker diarization, and audio event tagging.",
             inputs=[
@@ -210,7 +210,7 @@ class ElevenLabsVoiceSelector(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsVoiceSelector",
             display_name="ElevenLabs Voice Selector",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Select a predefined ElevenLabs voice for text-to-speech generation.",
             inputs=[
                 IO.Combo.Input(
@@ -239,7 +239,7 @@ class ElevenLabsTextToSpeech(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsTextToSpeech",
             display_name="ElevenLabs Text to Speech",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Convert text to speech.",
             inputs=[
                 IO.Custom(ELEVENLABS_VOICE).Input(
@@ -414,7 +414,7 @@ class ElevenLabsAudioIsolation(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsAudioIsolation",
             display_name="ElevenLabs Voice Isolation",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Remove background noise from audio, isolating vocals or speech.",
             inputs=[
                 IO.Audio.Input(
@@ -459,7 +459,7 @@ class ElevenLabsTextToSoundEffects(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsTextToSoundEffects",
             display_name="ElevenLabs Text to Sound Effects",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Generate sound effects from text descriptions.",
             inputs=[
                 IO.String.Input(
@@ -555,7 +555,7 @@ class ElevenLabsInstantVoiceClone(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsInstantVoiceClone",
             display_name="ElevenLabs Instant Voice Clone",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Create a cloned voice from audio samples. "
             "Provide 1-8 audio recordings of the voice to clone.",
             inputs=[
@@ -658,7 +658,7 @@ class ElevenLabsSpeechToSpeech(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsSpeechToSpeech",
             display_name="ElevenLabs Speech to Speech",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Transform speech from one voice to another while preserving the original content and emotion.",
             inputs=[
                 IO.Custom(ELEVENLABS_VOICE).Input(
@@ -793,7 +793,7 @@ class ElevenLabsTextToDialogue(IO.ComfyNode):
         return IO.Schema(
             node_id="ElevenLabsTextToDialogue",
             display_name="ElevenLabs Text to Dialogue",
-            category="api node/audio/ElevenLabs",
+            category="audio/partner/ElevenLabs",
             description="Generate multi-speaker dialogue from text. Each dialogue entry has its own text and voice.",
             inputs=[
                 IO.Float.Input(

comfy_api_nodes/nodes_gemini.py

@@ -300,7 +300,7 @@ class GeminiNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GeminiNode",
             display_name="Google Gemini",
-            category="api node/text/Gemini",
+            category="text/partner/Gemini",
             description="Generate text responses with Google's Gemini AI model. "
             "You can provide multiple types of inputs (text, images, audio, video) "
             "as context for generating more relevant and meaningful responses.",
@@ -541,7 +541,7 @@ class GeminiInputFiles(IO.ComfyNode):
         return IO.Schema(
             node_id="GeminiInputFiles",
             display_name="Gemini Input Files",
-            category="api node/text/Gemini",
+            category="text/partner/Gemini",
             description="Loads and prepares input files to include as inputs for Gemini LLM nodes. "
             "The files will be read by the Gemini model when generating a response. "
             "The contents of the text file count toward the token limit. "
@@ -598,7 +598,7 @@ class GeminiImage(IO.ComfyNode):
         return IO.Schema(
             node_id="GeminiImageNode",
             display_name="Nano Banana (Google Gemini Image)",
-            category="api node/image/Gemini",
+            category="image/partner/Gemini",
             description="Edit images synchronously via Google API.",
             inputs=[
                 IO.String.Input(
@@ -731,7 +731,7 @@ class GeminiImage2(IO.ComfyNode):
         return IO.Schema(
             node_id="GeminiImage2Node",
             display_name="Nano Banana Pro (Google Gemini Image)",
-            category="api node/image/Gemini",
+            category="image/partner/Gemini",
             description="Generate or edit images synchronously via Google Vertex API.",
             inputs=[
                 IO.String.Input(
@@ -869,7 +869,7 @@ class GeminiNanoBanana2(IO.ComfyNode):
         return IO.Schema(
             node_id="GeminiNanoBanana2",
             display_name="Nano Banana 2",
-            category="api node/image/Gemini",
+            category="image/partner/Gemini",
             description="Generate or edit images synchronously via Google Vertex API.",
             inputs=[
                 IO.String.Input(
@@ -1085,7 +1085,7 @@ class GeminiNanoBanana2V2(IO.ComfyNode):
         return IO.Schema(
             node_id="GeminiNanoBanana2V2",
             display_name="Nano Banana 2",
-            category="api node/image/Gemini",
+            category="image/partner/Gemini",
             description="Generate or edit images synchronously via Google Vertex API.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_grok.py

@@ -49,7 +49,7 @@ class GrokImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokImageNode",
             display_name="Grok Image",
-            category="api node/image/Grok",
+            category="image/partner/Grok",
             description="Generate images using Grok based on a text prompt",
             inputs=[
                 IO.Combo.Input(
@@ -58,7 +58,6 @@ class GrokImageNode(IO.ComfyNode):
                         "grok-imagine-image-quality",
                         "grok-imagine-image-pro",
                         "grok-imagine-image",
-                        "grok-imagine-image-beta",
                     ],
                 ),
                 IO.String.Input(
@@ -224,7 +223,7 @@ class GrokImageEditNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokImageEditNode",
             display_name="Grok Image Edit",
-            category="api node/image/Grok",
+            category="image/partner/Grok",
             description="Modify an existing image based on a text prompt",
             inputs=[
                 IO.Combo.Input(
@@ -233,7 +232,6 @@ class GrokImageEditNode(IO.ComfyNode):
                         "grok-imagine-image-quality",
                         "grok-imagine-image-pro",
                         "grok-imagine-image",
-                        "grok-imagine-image-beta",
                     ],
                 ),
                 IO.Image.Input("image", display_name="images"),
@@ -366,7 +364,7 @@ class GrokImageEditNodeV2(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokImageEditNodeV2",
             display_name="Grok Image Edit",
-            category="api node/image/Grok",
+            category="image/partner/Grok",
             description="Modify an existing image based on a text prompt",
             inputs=[
                 IO.String.Input(
@@ -503,10 +501,10 @@ class GrokVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokVideoNode",
             display_name="Grok Video",
-            category="api node/video/Grok",
+            category="video/partner/Grok",
             description="Generate video from a prompt or an image",
             inputs=[
-                IO.Combo.Input("model", options=["grok-imagine-video", "grok-imagine-video-beta"]),
+                IO.Combo.Input("model", options=["grok-imagine-video"]),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
@@ -576,8 +574,6 @@ class GrokVideoNode(IO.ComfyNode):
         seed: int,
         image: Input.Image | None = None,
     ) -> IO.NodeOutput:
-        if model == "grok-imagine-video-beta":
-            model = "grok-imagine-video"
         image_url = None
         if image is not None:
             if get_number_of_images(image) != 1:
@@ -615,10 +611,10 @@ class GrokVideoEditNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokVideoEditNode",
             display_name="Grok Video Edit",
-            category="api node/video/Grok",
+            category="video/partner/Grok",
             description="Edit an existing video based on a text prompt.",
             inputs=[
-                IO.Combo.Input("model", options=["grok-imagine-video", "grok-imagine-video-beta"]),
+                IO.Combo.Input("model", options=["grok-imagine-video"]),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
@@ -693,7 +689,7 @@ class GrokVideoReferenceNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokVideoReferenceNode",
             display_name="Grok Reference-to-Video",
-            category="api node/video/Grok",
+            category="video/partner/Grok",
             description="Generate video guided by reference images as style and content references.",
             inputs=[
                 IO.String.Input(
@@ -826,7 +822,7 @@ class GrokVideoExtendNode(IO.ComfyNode):
         return IO.Schema(
             node_id="GrokVideoExtendNode",
             display_name="Grok Video Extend",
-            category="api node/video/Grok",
+            category="video/partner/Grok",
             description="Extend an existing video with a seamless continuation based on a text prompt.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_hitpaw.py

@@ -71,7 +71,7 @@ class HitPawGeneralImageEnhance(IO.ComfyNode):
         return IO.Schema(
             node_id="HitPawGeneralImageEnhance",
             display_name="HitPaw General Image Enhance",
-            category="api node/image/HitPaw",
+            category="image/partner/HitPaw",
             description="Upscale low-resolution images to super-resolution, eliminate artifacts and noise. "
             f"Maximum output: {MAX_MP_GENERATIVE} megapixels.",
             inputs=[
@@ -201,7 +201,7 @@ class HitPawVideoEnhance(IO.ComfyNode):
         return IO.Schema(
             node_id="HitPawVideoEnhance",
             display_name="HitPaw Video Enhance",
-            category="api node/video/HitPaw",
+            category="video/partner/HitPaw",
             description="Upscale low-resolution videos to high resolution, eliminate artifacts and noise. "
             "Prices shown are per second of video.",
             inputs=[

comfy_api_nodes/nodes_hunyuan3d.py

@@ -123,7 +123,7 @@ class TencentTextToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TencentTextToModelNode",
             display_name="Hunyuan3D: Text to Model",
-            category="api node/3d/Tencent",
+            category="3d/partner/Tencent",
             essentials_category="3D",
             inputs=[
                 IO.Combo.Input(
@@ -242,7 +242,7 @@ class TencentImageToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TencentImageToModelNode",
             display_name="Hunyuan3D: Image(s) to Model",
-            category="api node/3d/Tencent",
+            category="3d/partner/Tencent",
             essentials_category="3D",
             inputs=[
                 IO.Combo.Input(
@@ -415,7 +415,7 @@ class TencentModelTo3DUVNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TencentModelTo3DUVNode",
             display_name="Hunyuan3D: Model to UV",
-            category="api node/3d/Tencent",
+            category="3d/partner/Tencent",
             description="Perform UV unfolding on a 3D model to generate UV texture. "
             "Input model must have less than 30000 faces.",
             inputs=[
@@ -505,7 +505,7 @@ class Tencent3DTextureEditNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Tencent3DTextureEditNode",
             display_name="Hunyuan3D: 3D Texture Edit",
-            category="api node/3d/Tencent",
+            category="3d/partner/Tencent",
             description="After inputting the 3D model, perform 3D model texture redrawing.",
             inputs=[
                 IO.MultiType.Input(
@@ -594,7 +594,7 @@ class Tencent3DPartNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Tencent3DPartNode",
             display_name="Hunyuan3D: 3D Part",
-            category="api node/3d/Tencent",
+            category="3d/partner/Tencent",
             description="Automatically perform component identification and generation based on the model structure.",
             inputs=[
                 IO.MultiType.Input(
@@ -666,7 +666,7 @@ class TencentSmartTopologyNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TencentSmartTopologyNode",
             display_name="Hunyuan3D: Smart Topology",
-            category="api node/3d/Tencent",
+            category="3d/partner/Tencent",
             description="Perform smart retopology on a 3D model. "
             "Supports GLB/OBJ formats; max 200MB; recommended for high-poly models.",
             inputs=[

comfy_api_nodes/nodes_ideogram.py

@@ -234,7 +234,7 @@ class IdeogramV1(IO.ComfyNode):
         return IO.Schema(
             node_id="IdeogramV1",
             display_name="Ideogram V1",
-            category="api node/image/Ideogram",
+            category="image/partner/Ideogram",
             description="Generates images using the Ideogram V1 model.",
             inputs=[
                 IO.String.Input(
@@ -360,7 +360,7 @@ class IdeogramV2(IO.ComfyNode):
         return IO.Schema(
             node_id="IdeogramV2",
             display_name="Ideogram V2",
-            category="api node/image/Ideogram",
+            category="image/partner/Ideogram",
             description="Generates images using the Ideogram V2 model.",
             inputs=[
                 IO.String.Input(
@@ -526,7 +526,7 @@ class IdeogramV3(IO.ComfyNode):
         return IO.Schema(
             node_id="IdeogramV3",
             display_name="Ideogram V3",
-            category="api node/image/Ideogram",
+            category="image/partner/Ideogram",
             description="Generates images using the Ideogram V3 model. "
                         "Supports both regular image generation from text prompts and image editing with mask.",
             inputs=[

comfy_api_nodes/nodes_kling.py

@@ -642,7 +642,7 @@ class KlingCameraControls(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingCameraControls",
             display_name="Kling Camera Controls",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Allows specifying configuration options for Kling Camera Controls and motion control effects.",
             inputs=[
                 IO.Combo.Input("camera_control_type", options=KlingCameraControlType),
@@ -762,7 +762,7 @@ class KlingTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingTextToVideoNode",
             display_name="Kling Text to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Kling Text to Video Node",
             inputs=[
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"),
@@ -849,7 +849,7 @@ class OmniProTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingOmniProTextToVideoNode",
             display_name="Kling 3.0 Omni Text to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Use text prompts to generate videos with the latest Kling model.",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]),
@@ -998,7 +998,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingOmniProFirstLastFrameNode",
             display_name="Kling 3.0 Omni First-Last-Frame to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Use a start frame, an optional end frame, or reference images with the latest Kling model.",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]),
@@ -1205,7 +1205,7 @@ class OmniProImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingOmniProImageToVideoNode",
             display_name="Kling 3.0 Omni Image to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Use up to 7 reference images to generate a video with the latest Kling model.",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]),
@@ -1374,7 +1374,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingOmniProVideoToVideoNode",
             display_name="Kling 3.0 Omni Video to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Use a video and up to 4 reference images to generate a video with the latest Kling model.",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]),
@@ -1485,7 +1485,7 @@ class OmniProEditVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingOmniProEditVideoNode",
             display_name="Kling 3.0 Omni Edit Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             essentials_category="Video Generation",
             description="Edit an existing video with the latest model from Kling.",
             inputs=[
@@ -1593,7 +1593,7 @@ class OmniProImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingOmniProImageNode",
             display_name="Kling 3.0 Omni Image",
-            category="api node/image/Kling",
+            category="image/partner/Kling",
             description="Create or edit images with the latest model from Kling.",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-image-o1"]),
@@ -1721,7 +1721,7 @@ class KlingCameraControlT2VNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingCameraControlT2VNode",
             display_name="Kling Text to Video (Camera Control)",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Transform text into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original text.",
             inputs=[
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"),
@@ -1783,7 +1783,7 @@ class KlingImage2VideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingImage2VideoNode",
             display_name="Kling Image(First Frame) to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             inputs=[
                 IO.Image.Input("start_frame", tooltip="The reference image used to generate the video."),
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"),
@@ -1882,7 +1882,7 @@ class KlingCameraControlI2VNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingCameraControlI2VNode",
             display_name="Kling Image to Video (Camera Control)",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Transform still images into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original image.",
             inputs=[
                 IO.Image.Input(
@@ -1953,7 +1953,7 @@ class KlingStartEndFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingStartEndFrameNode",
             display_name="Kling Start-End Frame to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Generate a video sequence that transitions between your provided start and end images. The node creates all frames in between, producing a smooth transformation from the first frame to the last.",
             inputs=[
                 IO.Image.Input(
@@ -2047,7 +2047,7 @@ class KlingVideoExtendNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingVideoExtendNode",
             display_name="Kling Video Extend",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Kling Video Extend Node. Extend videos made by other Kling nodes. The video_id is created by using other Kling Nodes.",
             inputs=[
                 IO.String.Input(
@@ -2128,7 +2128,7 @@ class KlingDualCharacterVideoEffectNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingDualCharacterVideoEffectNode",
             display_name="Kling Dual Character Video Effects",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Achieve different special effects when generating a video based on the effect_scene. First image will be positioned on left side, second on right side of the composite.",
             inputs=[
                 IO.Image.Input("image_left", tooltip="Left side image"),
@@ -2218,7 +2218,7 @@ class KlingSingleImageVideoEffectNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingSingleImageVideoEffectNode",
             display_name="Kling Video Effects",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Achieve different special effects when generating a video based on the effect_scene.",
             inputs=[
                 IO.Image.Input(
@@ -2291,7 +2291,7 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingLipSyncAudioToVideoNode",
             display_name="Kling Lip Sync Video with Audio",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             essentials_category="Video Generation",
             description="Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file. When using, ensure that the audio contains clearly distinguishable vocals and that the video contains a distinct face. The audio file should not be larger than 5MB. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length.",
             inputs=[
@@ -2343,7 +2343,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingLipSyncTextToVideoNode",
             display_name="Kling Lip Sync Video with Text",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Kling Lip Sync Text to Video Node. Syncs mouth movements in a video file to a text prompt. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length.",
             inputs=[
                 IO.Video.Input("video"),
@@ -2411,7 +2411,7 @@ class KlingVirtualTryOnNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingVirtualTryOnNode",
             display_name="Kling Virtual Try On",
-            category="api node/image/Kling",
+            category="image/partner/Kling",
             description="Kling Virtual Try On Node. Input a human image and a cloth image to try on the cloth on the human. You can merge multiple clothing item pictures into one image with a white background.",
             inputs=[
                 IO.Image.Input("human_image"),
@@ -2478,7 +2478,7 @@ class KlingImageGenerationNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingImageGenerationNode",
             display_name="Kling 3.0 Image",
-            category="api node/image/Kling",
+            category="image/partner/Kling",
             description="Kling Image Generation Node. Generate an image from a text prompt with an optional reference image.",
             inputs=[
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"),
@@ -2615,7 +2615,7 @@ class TextToVideoWithAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingTextToVideoWithAudio",
             display_name="Kling 2.6 Text to Video with Audio",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v2-6"]),
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."),
@@ -2683,7 +2683,7 @@ class ImageToVideoWithAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingImageToVideoWithAudio",
             display_name="Kling 2.6 Image(First Frame) to Video with Audio",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             inputs=[
                 IO.Combo.Input("model_name", options=["kling-v2-6"]),
                 IO.Image.Input("start_frame"),
@@ -2753,7 +2753,7 @@ class MotionControl(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingMotionControl",
             display_name="Kling Motion Control",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             inputs=[
                 IO.String.Input("prompt", multiline=True),
                 IO.Image.Input("reference_image"),
@@ -2854,7 +2854,7 @@ class KlingVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingVideoNode",
             display_name="Kling 3.0 Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Generate videos with Kling V3. "
             "Supports text-to-video and image-to-video with optional storyboard multi-prompt and audio generation.",
             inputs=[
@@ -3077,7 +3077,7 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingFirstLastFrameNode",
             display_name="Kling 3.0 First-Last-Frame to Video",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Generate videos with Kling V3 using first and last frames.",
             inputs=[
                 IO.String.Input("prompt", multiline=True, default=""),
@@ -3202,7 +3202,7 @@ class KlingAvatarNode(IO.ComfyNode):
         return IO.Schema(
             node_id="KlingAvatarNode",
             display_name="Kling Avatar 2.0",
-            category="api node/video/Kling",
+            category="video/partner/Kling",
             description="Generate broadcast-style digital human videos from a single photo and an audio file.",
             inputs=[
                 IO.Image.Input(

comfy_api_nodes/nodes_krea.py

@@ -0,0 +1,290 @@
+"""Krea image-generation nodes."""
+
+import re
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.krea import (
+    KreaAssetResponse,
+    KreaGenerateImageRequest,
+    KreaImageStyleReference,
+    KreaJob,
+    KreaMoodboard,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    download_url_to_image_tensor,
+    poll_op,
+    sync_op,
+    tensor_to_bytesio,
+    validate_string,
+)
+
+
+class KreaIO:
+    STYLE_REF = "KREA_STYLE_REF"
+
+
+async def _upload_image_to_krea_assets(cls: type[IO.ComfyNode], image: Input.Image) -> str:
+    """Upload an image to Krea's /assets endpoint and return the Krea-hosted image URL."""
+    img_io = tensor_to_bytesio(image, total_pixels=2048 * 2048, mime_type="image/png")
+    response = await sync_op(
+        cls,
+        endpoint=ApiEndpoint(path="/proxy/krea/assets", method="POST"),
+        response_model=KreaAssetResponse,
+        files=[("file", (img_io.name, img_io, "image/png"))],
+        content_type="multipart/form-data",
+        max_retries=1,
+        wait_label="Uploading reference",
+    )
+    return response.image_url
+
+
+_MODEL_MEDIUM = "Krea 2 Medium"
+_MODEL_LARGE = "Krea 2 Large"
+_MODEL_ENDPOINTS: dict[str, str] = {
+    _MODEL_MEDIUM: "/proxy/krea/generate/image/krea/krea-2/medium",
+    _MODEL_LARGE: "/proxy/krea/generate/image/krea/krea-2/large",
+}
+
+_ASPECT_RATIOS = ["1:1", "4:3", "3:2", "16:9", "2.35:1", "4:5", "2:3", "9:16"]
+_RESOLUTIONS = ["1K"]
+_CREATIVITY_LEVELS = ["raw", "low", "medium", "high"]
+_KREA_QUEUED_STATUSES = ["backlogged", "queued", "scheduled"]
+
+_UUID_RE = re.compile(r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$")
+
+
+def _krea_model_inputs() -> list:
+    """Nested inputs shared by both Krea 2 Medium and Large under the DynamicCombo."""
+    return [
+        IO.Combo.Input(
+            "aspect_ratio",
+            options=_ASPECT_RATIOS,
+            tooltip="Output aspect ratio.",
+        ),
+        IO.Combo.Input(
+            "resolution",
+            options=_RESOLUTIONS,
+            tooltip="Resolution scale.",
+        ),
+        IO.Combo.Input(
+            "creativity",
+            options=_CREATIVITY_LEVELS,
+            default="medium",
+            tooltip="Prompt interpretation strength: raw stays closest to the prompt; high is most creative.",
+        ),
+        IO.String.Input(
+            "moodboard_id",
+            default="",
+            tooltip="Optional Krea moodboard UUID (e.g. from the Krea website). "
+            "Leave empty to disable. Only one moodboard is supported per request.",
+            optional=True,
+        ),
+        IO.Float.Input(
+            "moodboard_strength",
+            default=0.35,
+            min=-0.5,
+            max=1.5,
+            step=0.05,
+            tooltip="Moodboard influence; ignored when moodboard_id is empty.",
+            optional=True,
+        ),
+        IO.Custom(KreaIO.STYLE_REF).Input(
+            "style_reference",
+            optional=True,
+            tooltip="Optional chain of style references (max 10) from Krea 2 Style Reference nodes.",
+        ),
+    ]
+
+
+class Krea2ImageNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="Krea2ImageNode",
+            display_name="Krea 2 Image",
+            category="image/partner/Krea",
+            description=(
+                "Generate images via Krea 2 — pick Medium (expressive illustrations) or "
+                "Large (expressive photorealism). Supports an optional moodboard and up "
+                "to 10 chained image style references."
+            ),
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text prompt for the image.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(_MODEL_MEDIUM, _krea_model_inputs()),
+                        IO.DynamicCombo.Option(_MODEL_LARGE, _krea_model_inputs()),
+                    ],
+                    tooltip="Krea 2 Medium is best for expressive illustrations; "
+                    "Krea 2 Large is best for expressive photorealism.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    control_after_generate=True,
+                    tooltip="Random seed for reproducibility.",
+                ),
+            ],
+            outputs=[IO.Image.Output()],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(
+                    widgets=["model", "model.moodboard_id"],
+                    inputs=["model.style_reference"],
+                ),
+                expr="""
+                (
+                  $isLarge := widgets.model = "krea 2 large";
+                  $hasMoodboard := $length($lookup(widgets, "model.moodboard_id")) > 0;
+                  $hasStyle := $lookup(inputs, "model.style_reference").connected;
+                  $usd := $hasMoodboard
+                    ? ($isLarge ? 0.07 : 0.04)
+                    : ($hasStyle
+                        ? ($isLarge ? 0.065 : 0.035)
+                        : ($isLarge ? 0.06 : 0.03));
+                  {"type":"usd","usd": $usd}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        model: dict,
+        seed: int,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, strip_whitespace=False, min_length=1)
+
+        model_choice = model["model"]
+        endpoint_path = _MODEL_ENDPOINTS.get(model_choice)
+        if endpoint_path is None:
+            raise ValueError(f"Unknown Krea 2 model: {model_choice!r}")
+
+        moodboards: list[KreaMoodboard] | None = None
+        mb_id = (model.get("moodboard_id") or "").strip()
+        if mb_id:
+            if not _UUID_RE.match(mb_id):
+                raise ValueError(f"moodboard_id must be a UUID (received {mb_id!r}); copy it from the Krea website.")
+            mb_strength = model.get("moodboard_strength")
+            moodboards = [KreaMoodboard(id=mb_id, strength=0.35 if mb_strength is None else float(mb_strength))]
+
+        style_reference = model.get("style_reference")
+        image_style_references: list[KreaImageStyleReference] | None = None
+        if style_reference:
+            if len(style_reference) > 10:
+                raise ValueError(f"Krea 2 accepts at most 10 image_style_references; received {len(style_reference)}.")
+            image_style_references = [
+                KreaImageStyleReference(url=ref["url"], strength=float(ref["strength"])) for ref in style_reference
+            ]
+        initial = await sync_op(
+            cls,
+            ApiEndpoint(path=endpoint_path, method="POST"),
+            response_model=KreaJob,
+            data=KreaGenerateImageRequest(
+                prompt=prompt,
+                aspect_ratio=model["aspect_ratio"],
+                resolution=model["resolution"],
+                seed=seed,
+                creativity=model["creativity"],
+                moodboards=moodboards,
+                image_style_references=image_style_references,
+            ),
+        )
+        job = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/krea/jobs/{initial.job_id}", method="GET"),
+            response_model=KreaJob,
+            status_extractor=lambda r: r.status,
+            queued_statuses=_KREA_QUEUED_STATUSES,
+        )
+        if not job.result or not job.result.urls:
+            raise RuntimeError(f"Krea 2 job {job.job_id} completed without any image URLs.")
+        image = await download_url_to_image_tensor(job.result.urls[0])
+        return IO.NodeOutput(image)
+
+
+class Krea2StyleReferenceNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="Krea2StyleReferenceNode",
+            display_name="Krea 2 Style Reference",
+            category="image/partner/Krea",
+            description=(
+                "Add an image style reference to a Krea 2 generation. Chain multiple Krea 2 "
+                "Style Reference nodes (max 10) and feed the final `style_reference` output "
+                "into Krea 2 Image. Each image is uploaded to ComfyAPI storage and passed as URL."
+            ),
+            inputs=[
+                IO.Image.Input(
+                    "image",
+                    tooltip="Reference image whose style influences the generation.",
+                ),
+                IO.Float.Input(
+                    "strength",
+                    default=1.0,
+                    min=-2.0,
+                    max=2.0,
+                    step=0.05,
+                    tooltip="Reference strength; negative values invert the style influence.",
+                ),
+                IO.Custom(KreaIO.STYLE_REF).Input(
+                    "style_reference",
+                    optional=True,
+                    tooltip="Optional incoming chain of style references; this node appends one more.",
+                ),
+            ],
+            outputs=[IO.Custom(KreaIO.STYLE_REF).Output(display_name="style_reference")],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        strength: float,
+        style_reference: list[dict] | None = None,
+    ) -> IO.NodeOutput:
+        chain: list[dict] = list(style_reference) if style_reference else []
+        if len(chain) >= 10:
+            raise ValueError("Krea 2 accepts at most 10 image_style_references in one generation.")
+        url = await _upload_image_to_krea_assets(cls, image)
+        chain.append({"url": url, "strength": float(strength)})
+        return IO.NodeOutput(chain)
+
+
+class KreaExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            Krea2ImageNode,
+            Krea2StyleReferenceNode,
+        ]
+
+
+async def comfy_entrypoint() -> KreaExtension:
+    return KreaExtension()

comfy_api_nodes/nodes_ltxv.py

@@ -50,7 +50,7 @@ class TextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LtxvApiTextToVideo",
             display_name="LTXV Text To Video",
-            category="api node/video/LTXV",
+            category="video/partner/LTXV",
             description="Professional-quality videos with customizable duration and resolution.",
             inputs=[
                 IO.Combo.Input("model", options=list(MODELS_MAP.keys())),
@@ -127,7 +127,7 @@ class ImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LtxvApiImageToVideo",
             display_name="LTXV Image To Video",
-            category="api node/video/LTXV",
+            category="video/partner/LTXV",
             description="Professional-quality videos with customizable duration and resolution based on start image.",
             inputs=[
                 IO.Image.Input("image", tooltip="First frame to be used for the video."),

comfy_api_nodes/nodes_luma.py

@@ -46,7 +46,7 @@ class LumaReferenceNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaReferenceNode",
             display_name="Luma Reference",
-            category="api node/image/Luma",
+            category="image/partner/Luma",
             description="Holds an image and weight for use with Luma Generate Image node.",
             inputs=[
                 IO.Image.Input(
@@ -85,7 +85,7 @@ class LumaConceptsNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaConceptsNode",
             display_name="Luma Concepts",
-            category="api node/video/Luma",
+            category="video/partner/Luma",
             description="Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes.",
             inputs=[
                 IO.Combo.Input(
@@ -134,7 +134,7 @@ class LumaImageGenerationNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaImageNode",
             display_name="Luma Text to Image",
-            category="api node/image/Luma",
+            category="image/partner/Luma",
             description="Generates images synchronously based on prompt and aspect ratio.",
             inputs=[
                 IO.String.Input(
@@ -278,7 +278,7 @@ class LumaImageModifyNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaImageModifyNode",
             display_name="Luma Image to Image",
-            category="api node/image/Luma",
+            category="image/partner/Luma",
             description="Modifies images synchronously based on prompt and aspect ratio.",
             inputs=[
                 IO.Image.Input(
@@ -371,7 +371,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaVideoNode",
             display_name="Luma Text to Video",
-            category="api node/video/Luma",
+            category="video/partner/Luma",
             description="Generates videos synchronously based on prompt and output_size.",
             inputs=[
                 IO.String.Input(
@@ -472,7 +472,7 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaImageToVideoNode",
             display_name="Luma Image to Video",
-            category="api node/video/Luma",
+            category="video/partner/Luma",
             description="Generates videos synchronously based on prompt, input images, and output_size.",
             inputs=[
                 IO.String.Input(
@@ -724,7 +724,7 @@ class LumaImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaImageNode2",
             display_name="Luma UNI-1 Image",
-            category="api node/image/Luma",
+            category="image/partner/Luma",
             description="Generate images from text using the Luma UNI-1 model.",
             inputs=[
                 IO.String.Input(
@@ -853,7 +853,7 @@ class LumaImageEditNode(IO.ComfyNode):
         return IO.Schema(
             node_id="LumaImageEditNode2",
             display_name="Luma UNI-1 Image Edit",
-            category="api node/image/Luma",
+            category="image/partner/Luma",
             description="Edit an existing image with a text prompt using the Luma UNI-1 model.",
             inputs=[
                 IO.Image.Input(

comfy_api_nodes/nodes_magnific.py

@@ -61,7 +61,7 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MagnificImageUpscalerCreativeNode",
             display_name="Magnific Image Upscale (Creative)",
-            category="api node/image/Magnific",
+            category="image/partner/Magnific",
             description="Prompt‑guided enhancement, stylization, and 2x/4x/8x/16x upscaling. "
             "Maximum output: 25.3 megapixels.",
             inputs=[
@@ -240,7 +240,7 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
         return IO.Schema(
             node_id="MagnificImageUpscalerPreciseV2Node",
             display_name="Magnific Image Upscale (Precise V2)",
-            category="api node/image/Magnific",
+            category="image/partner/Magnific",
             description="High-fidelity upscaling with fine control over sharpness, grain, and detail. "
             "Maximum output: 10060×10060 pixels.",
             inputs=[
@@ -400,7 +400,7 @@ class MagnificImageStyleTransferNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MagnificImageStyleTransferNode",
             display_name="Magnific Image Style Transfer",
-            category="api node/image/Magnific",
+            category="image/partner/Magnific",
             description="Transfer the style from a reference image to your input image.",
             inputs=[
                 IO.Image.Input("image", tooltip="The image to apply style transfer to."),
@@ -549,7 +549,7 @@ class MagnificImageRelightNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MagnificImageRelightNode",
             display_name="Magnific Image Relight",
-            category="api node/image/Magnific",
+            category="image/partner/Magnific",
             description="Relight an image with lighting adjustments and optional reference-based light transfer.",
             inputs=[
                 IO.Image.Input("image", tooltip="The image to relight."),
@@ -789,7 +789,7 @@ class MagnificImageSkinEnhancerNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MagnificImageSkinEnhancerNode",
             display_name="Magnific Image Skin Enhancer",
-            category="api node/image/Magnific",
+            category="image/partner/Magnific",
             description="Skin enhancement for portraits with multiple processing modes.",
             inputs=[
                 IO.Image.Input("image", tooltip="The portrait image to enhance."),

comfy_api_nodes/nodes_meshy.py

@@ -33,7 +33,7 @@ class MeshyTextToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyTextToModelNode",
             display_name="Meshy: Text to Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             inputs=[
                 IO.Combo.Input("model", options=["latest"]),
                 IO.String.Input("prompt", multiline=True, default=""),
@@ -145,7 +145,7 @@ class MeshyRefineNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyRefineNode",
             display_name="Meshy: Refine Draft Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             description="Refine a previously created draft model.",
             inputs=[
                 IO.Combo.Input("model", options=["latest"]),
@@ -240,7 +240,7 @@ class MeshyImageToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyImageToModelNode",
             display_name="Meshy: Image to Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             inputs=[
                 IO.Combo.Input("model", options=["latest"]),
                 IO.Image.Input("image"),
@@ -405,7 +405,7 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyMultiImageToModelNode",
             display_name="Meshy: Multi-Image to Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             inputs=[
                 IO.Combo.Input("model", options=["latest"]),
                 IO.Autogrow.Input(
@@ -575,7 +575,7 @@ class MeshyRigModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyRigModelNode",
             display_name="Meshy: Rig Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             description="Provides a rigged character in standard formats. "
             "Auto-rigging is currently not suitable for untextured meshes, non-humanoid assets, "
             "or humanoid assets with unclear limb and body structure.",
@@ -656,7 +656,7 @@ class MeshyAnimateModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyAnimateModelNode",
             display_name="Meshy: Animate Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             description="Apply a specific animation action to a previously rigged character.",
             inputs=[
                 IO.Custom("MESHY_RIGGED_TASK_ID").Input("rig_task_id"),
@@ -722,7 +722,7 @@ class MeshyTextureNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MeshyTextureNode",
             display_name="Meshy: Texture Model",
-            category="api node/3d/Meshy",
+            category="3d/partner/Meshy",
             inputs=[
                 IO.Combo.Input("model", options=["latest"]),
                 IO.Custom("MESHY_TASK_ID").Input("meshy_task_id"),

comfy_api_nodes/nodes_minimax.py

@@ -101,7 +101,7 @@ class MinimaxTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MinimaxTextToVideoNode",
             display_name="MiniMax Text to Video",
-            category="api node/video/MiniMax",
+            category="video/partner/MiniMax",
             description="Generates videos synchronously based on a prompt, and optional parameters.",
             inputs=[
                 IO.String.Input(
@@ -163,7 +163,7 @@ class MinimaxImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MinimaxImageToVideoNode",
             display_name="MiniMax Image to Video",
-            category="api node/video/MiniMax",
+            category="video/partner/MiniMax",
             description="Generates videos synchronously based on an image and prompt, and optional parameters.",
             inputs=[
                 IO.Image.Input(
@@ -230,7 +230,7 @@ class MinimaxSubjectToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MinimaxSubjectToVideoNode",
             display_name="MiniMax Subject to Video",
-            category="api node/video/MiniMax",
+            category="video/partner/MiniMax",
             description="Generates videos synchronously based on an image and prompt, and optional parameters.",
             inputs=[
                 IO.Image.Input(
@@ -294,7 +294,7 @@ class MinimaxHailuoVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="MinimaxHailuoVideoNode",
             display_name="MiniMax Hailuo Video",
-            category="api node/video/MiniMax",
+            category="video/partner/MiniMax",
             description="Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_openai.py

@@ -99,7 +99,7 @@ class OpenAIDalle2(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIDalle2",
             display_name="OpenAI DALL·E 2",
-            category="api node/image/OpenAI",
+            category="image/partner/OpenAI",
             description="Generates images synchronously via OpenAI's DALL·E 2 endpoint.",
             inputs=[
                 IO.String.Input(
@@ -249,7 +249,7 @@ class OpenAIDalle3(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIDalle3",
             display_name="OpenAI DALL·E 3",
-            category="api node/image/OpenAI",
+            category="image/partner/OpenAI",
             description="Generates images synchronously via OpenAI's DALL·E 3 endpoint.",
             inputs=[
                 IO.String.Input(
@@ -371,7 +371,7 @@ class OpenAIGPTImage1(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIGPTImage1",
             display_name="OpenAI GPT Image 2",
-            category="api node/image/OpenAI",
+            category="image/partner/OpenAI",
             description="Generates images synchronously via OpenAI's GPT Image endpoint.",
             is_deprecated=True,
             inputs=[
@@ -695,7 +695,7 @@ class OpenAIGPTImageNodeV2(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIGPTImageNodeV2",
             display_name="OpenAI GPT Image 2",
-            category="api node/image/OpenAI",
+            category="image/partner/OpenAI",
             description="Generates images via OpenAI's GPT Image endpoint.",
             inputs=[
                 IO.String.Input(
@@ -962,7 +962,7 @@ class OpenAIChatNode(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIChatNode",
             display_name="OpenAI ChatGPT",
-            category="api node/text/OpenAI",
+            category="text/partner/OpenAI",
             essentials_category="Text Generation",
             description="Generate text responses from an OpenAI model.",
             inputs=[
@@ -1201,7 +1201,7 @@ class OpenAIInputFiles(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIInputFiles",
             display_name="OpenAI ChatGPT Input Files",
-            category="api node/text/OpenAI",
+            category="text/partner/OpenAI",
             description="Loads and prepares input files (text, pdf, etc.) to include as inputs for the OpenAI Chat Node. The files will be read by the OpenAI model when generating a response. 🛈 TIP: Can be chained together with other OpenAI Input File nodes.",
             inputs=[
                 IO.Combo.Input(
@@ -1248,7 +1248,7 @@ class OpenAIChatConfig(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIChatConfig",
             display_name="OpenAI ChatGPT Advanced Options",
-            category="api node/text/OpenAI",
+            category="text/partner/OpenAI",
             description="Allows specifying advanced configuration options for the OpenAI Chat Nodes.",
             inputs=[
                 IO.Combo.Input(

comfy_api_nodes/nodes_openrouter.py

@@ -265,7 +265,7 @@ class OpenRouterLLMNode(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenRouterLLMNode",
             display_name="OpenRouter LLM",
-            category="api node/text/OpenRouter",
+            category="text/partner/OpenRouter",
             essentials_category="Text Generation",
             description=(
                 "Generate text responses through OpenRouter. Routes to a curated set of popular "

comfy_api_nodes/nodes_pixverse.py

@@ -53,7 +53,7 @@ class PixverseTemplateNode(IO.ComfyNode):
         return IO.Schema(
             node_id="PixverseTemplateNode",
             display_name="PixVerse Template",
-            category="api node/video/PixVerse",
+            category="video/partner/PixVerse",
             inputs=[
                 IO.Combo.Input("template", options=list(pixverse_templates.keys())),
             ],
@@ -74,7 +74,7 @@ class PixverseTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="PixverseTextToVideoNode",
             display_name="PixVerse Text to Video",
-            category="api node/video/PixVerse",
+            category="video/partner/PixVerse",
             description="Generates videos based on prompt and output_size.",
             inputs=[
                 IO.String.Input(
@@ -192,7 +192,7 @@ class PixverseImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="PixverseImageToVideoNode",
             display_name="PixVerse Image to Video",
-            category="api node/video/PixVerse",
+            category="video/partner/PixVerse",
             description="Generates videos based on prompt and output_size.",
             inputs=[
                 IO.Image.Input("image"),
@@ -310,7 +310,7 @@ class PixverseTransitionVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="PixverseTransitionVideoNode",
             display_name="PixVerse Transition Video",
-            category="api node/video/PixVerse",
+            category="video/partner/PixVerse",
             description="Generates videos based on prompt and output_size.",
             inputs=[
                 IO.Image.Input("first_frame"),

comfy_api_nodes/nodes_quiver.py

@@ -62,7 +62,7 @@ class QuiverTextToSVGNode(IO.ComfyNode):
         return IO.Schema(
             node_id="QuiverTextToSVGNode",
             display_name="Quiver Text to SVG",
-            category="api node/image/Quiver",
+            category="image/partner/Quiver",
             description="Generate an SVG from a text prompt using Quiver AI.",
             inputs=[
                 IO.String.Input(
@@ -177,7 +177,7 @@ class QuiverImageToSVGNode(IO.ComfyNode):
         return IO.Schema(
             node_id="QuiverImageToSVGNode",
             display_name="Quiver Image to SVG",
-            category="api node/image/Quiver",
+            category="image/partner/Quiver",
             description="Vectorize a raster image into SVG using Quiver AI.",
             inputs=[
                 IO.Image.Input(

comfy_api_nodes/nodes_recraft.py

@@ -178,7 +178,7 @@ class RecraftColorRGBNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftColorRGB",
             display_name="Recraft Color RGB",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Create Recraft Color by choosing specific RGB values.",
             inputs=[
                 IO.Int.Input("r", default=0, min=0, max=255, tooltip="Red value of color."),
@@ -204,7 +204,7 @@ class RecraftControlsNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftControls",
             display_name="Recraft Controls",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Create Recraft Controls for customizing Recraft generation.",
             inputs=[
                 IO.Custom(RecraftIO.COLOR).Input("colors", optional=True),
@@ -228,7 +228,7 @@ class RecraftStyleV3RealisticImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftStyleV3RealisticImage",
             display_name="Recraft Style - Realistic Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Select realistic_image style and optional substyle.",
             inputs=[
                 IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)),
@@ -253,7 +253,7 @@ class RecraftStyleV3DigitalIllustrationNode(RecraftStyleV3RealisticImageNode):
         return IO.Schema(
             node_id="RecraftStyleV3DigitalIllustration",
             display_name="Recraft Style - Digital Illustration",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Select realistic_image style and optional substyle.",
             inputs=[
                 IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)),
@@ -272,7 +272,7 @@ class RecraftStyleV3VectorIllustrationNode(RecraftStyleV3RealisticImageNode):
         return IO.Schema(
             node_id="RecraftStyleV3VectorIllustrationNode",
             display_name="Recraft Style - Realistic Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Select realistic_image style and optional substyle.",
             inputs=[
                 IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)),
@@ -291,7 +291,7 @@ class RecraftStyleV3LogoRasterNode(RecraftStyleV3RealisticImageNode):
         return IO.Schema(
             node_id="RecraftStyleV3LogoRaster",
             display_name="Recraft Style - Logo Raster",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Select realistic_image style and optional substyle.",
             inputs=[
                 IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE, include_none=False)),
@@ -308,7 +308,7 @@ class RecraftStyleInfiniteStyleLibrary(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftStyleV3InfiniteStyleLibrary",
             display_name="Recraft Style - Infinite Style Library",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Choose style based on preexisting UUID from Recraft's Infinite Style Library.",
             inputs=[
                 IO.String.Input("style_id", default="", tooltip="UUID of style from Infinite Style Library."),
@@ -331,7 +331,7 @@ class RecraftCreateStyleNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftCreateStyleNode",
             display_name="Recraft Create Style",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Create a custom style from reference images. "
             "Upload 1-5 images to use as style references. "
             "Total size of all images is limited to 5 MB.",
@@ -400,7 +400,7 @@ class RecraftTextToImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftTextToImageNode",
             display_name="Recraft Text to Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Generates images synchronously based on prompt and resolution.",
             inputs=[
                 IO.String.Input("prompt", multiline=True, default="", tooltip="Prompt for the image generation."),
@@ -512,7 +512,7 @@ class RecraftImageToImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftImageToImageNode",
             display_name="Recraft Image to Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Modify image based on prompt and strength.",
             inputs=[
                 IO.Image.Input("image"),
@@ -630,7 +630,7 @@ class RecraftImageInpaintingNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftImageInpaintingNode",
             display_name="Recraft Image Inpainting",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Modify image based on prompt and mask.",
             inputs=[
                 IO.Image.Input("image"),
@@ -732,7 +732,7 @@ class RecraftTextToVectorNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftTextToVectorNode",
             display_name="Recraft Text to Vector",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Generates SVG synchronously based on prompt and resolution.",
             inputs=[
                 IO.String.Input("prompt", default="", tooltip="Prompt for the image generation.", multiline=True),
@@ -832,7 +832,7 @@ class RecraftVectorizeImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftVectorizeImageNode",
             display_name="Recraft Vectorize Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             essentials_category="Image Tools",
             description="Generates SVG synchronously from an input image.",
             inputs=[
@@ -876,7 +876,7 @@ class RecraftReplaceBackgroundNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftReplaceBackgroundNode",
             display_name="Recraft Replace Background",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Replace background on image, based on provided prompt.",
             inputs=[
                 IO.Image.Input("image"),
@@ -963,7 +963,7 @@ class RecraftRemoveBackgroundNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftRemoveBackgroundNode",
             display_name="Recraft Remove Background",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             essentials_category="Image Tools",
             description="Remove background from image, and return processed image and mask.",
             inputs=[
@@ -1012,7 +1012,7 @@ class RecraftCrispUpscaleNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftCrispUpscaleNode",
             display_name="Recraft Crisp Upscale Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Upscale image synchronously.\n"
             "Enhances a given raster image using ‘crisp upscale’ tool, "
             "increasing image resolution, making the image sharper and cleaner.",
@@ -1058,7 +1058,7 @@ class RecraftCreativeUpscaleNode(RecraftCrispUpscaleNode):
         return IO.Schema(
             node_id="RecraftCreativeUpscaleNode",
             display_name="Recraft Creative Upscale Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Upscale image synchronously.\n"
             "Enhances a given raster image using ‘creative upscale’ tool, "
             "boosting resolution with a focus on refining small details and faces.",
@@ -1086,7 +1086,7 @@ class RecraftV4TextToImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftV4TextToImageNode",
             display_name="Recraft V4 Text to Image",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Generates images using Recraft V4 or V4 Pro models.",
             inputs=[
                 IO.String.Input(
@@ -1210,7 +1210,7 @@ class RecraftV4TextToVectorNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RecraftV4TextToVectorNode",
             display_name="Recraft V4 Text to Vector",
-            category="api node/image/Recraft",
+            category="image/partner/Recraft",
             description="Generates SVG using Recraft V4 or V4 Pro models.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_reve.py

@@ -109,7 +109,7 @@ class ReveImageCreateNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ReveImageCreateNode",
             display_name="Reve Image Create",
-            category="api node/image/Reve",
+            category="image/partner/Reve",
             description="Generate images from text descriptions using Reve.",
             inputs=[
                 IO.String.Input(
@@ -200,7 +200,7 @@ class ReveImageEditNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ReveImageEditNode",
             display_name="Reve Image Edit",
-            category="api node/image/Reve",
+            category="image/partner/Reve",
             description="Edit images using natural language instructions with Reve.",
             inputs=[
                 IO.Image.Input("image", tooltip="The image to edit."),
@@ -300,7 +300,7 @@ class ReveImageRemixNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ReveImageRemixNode",
             display_name="Reve Image Remix",
-            category="api node/image/Reve",
+            category="image/partner/Reve",
             description="Combine reference images with text prompts to create new images using Reve.",
             inputs=[
                 IO.Autogrow.Input(

comfy_api_nodes/nodes_rodin.py

@@ -230,7 +230,7 @@ class Rodin3D_Regular(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Regular",
             display_name="Rodin 3D Generate - Regular Generate",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("Images"),
@@ -289,7 +289,7 @@ class Rodin3D_Detail(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Detail",
             display_name="Rodin 3D Generate - Detail Generate",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("Images"),
@@ -348,7 +348,7 @@ class Rodin3D_Smooth(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Smooth",
             display_name="Rodin 3D Generate - Smooth Generate",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("Images"),
@@ -406,7 +406,7 @@ class Rodin3D_Sketch(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Sketch",
             display_name="Rodin 3D Generate - Sketch Generate",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("Images"),
@@ -468,7 +468,7 @@ class Rodin3D_Gen2(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Gen2",
             display_name="Rodin 3D Generate - Gen-2 Generate",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("Images"),
@@ -941,7 +941,7 @@ class Rodin3D_Gen25_Image(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Gen25_Image",
             display_name="Rodin 3D Gen-2.5 - Image to 3D",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=(
                 "Generate a 3D model from 1-5 reference images via Rodin Gen-2.5. "
                 "Pick a mode (Fast / Regular / Extreme-High) to tune quality vs. cost."
@@ -1035,7 +1035,7 @@ class Rodin3D_Gen25_Text(IO.ComfyNode):
         return IO.Schema(
             node_id="Rodin3D_Gen25_Text",
             display_name="Rodin 3D Gen-2.5 - Text to 3D",
-            category="api node/3d/Rodin",
+            category="3d/partner/Rodin",
             description=(
                 "Generate a 3D model from a text prompt via Rodin Gen-2.5. "
                 "Pick a mode (Fast / Regular / Extreme-High) to tune quality vs. cost."

comfy_api_nodes/nodes_runway.py

@@ -140,7 +140,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode):
         return IO.Schema(
             node_id="RunwayImageToVideoNodeGen3a",
             display_name="Runway Image to Video (Gen3a Turbo)",
-            category="api node/video/Runway",
+            category="video/partner/Runway",
             description="Generate a video from a single starting frame using Gen3a Turbo model. "
             "Before diving in, review these best practices to ensure that "
             "your input selections will set your generation up for success: "
@@ -234,7 +234,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode):
         return IO.Schema(
             node_id="RunwayImageToVideoNodeGen4",
             display_name="Runway Image to Video (Gen4 Turbo)",
-            category="api node/video/Runway",
+            category="video/partner/Runway",
             description="Generate a video from a single starting frame using Gen4 Turbo model. "
             "Before diving in, review these best practices to ensure that "
             "your input selections will set your generation up for success: "
@@ -329,7 +329,7 @@ class RunwayFirstLastFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RunwayFirstLastFrameNode",
             display_name="Runway First-Last-Frame to Video",
-            category="api node/video/Runway",
+            category="video/partner/Runway",
             description="Upload first and last keyframes, draft a prompt, and generate a video. "
             "More complex transitions, such as cases where the Last frame is completely different "
             "from the First frame, may benefit from the longer 10s duration. "
@@ -440,7 +440,7 @@ class RunwayTextToImageNode(IO.ComfyNode):
         return IO.Schema(
             node_id="RunwayTextToImageNode",
             display_name="Runway Text to Image",
-            category="api node/image/Runway",
+            category="image/partner/Runway",
             description="Generate an image from a text prompt using Runway's Gen 4 model. "
             "You can also include reference image to guide the generation.",
             inputs=[

comfy_api_nodes/nodes_sonilo.py

@@ -34,7 +34,7 @@ class SoniloVideoToMusic(IO.ComfyNode):
         return IO.Schema(
             node_id="SoniloVideoToMusic",
             display_name="Sonilo Video to Music",
-            category="api node/audio/Sonilo",
+            category="audio/partner/Sonilo",
             description="Generate music from video content using Sonilo's AI model. "
             "Analyzes the video and creates matching music.",
             inputs=[
@@ -99,7 +99,7 @@ class SoniloTextToMusic(IO.ComfyNode):
         return IO.Schema(
             node_id="SoniloTextToMusic",
             display_name="Sonilo Text to Music",
-            category="api node/audio/Sonilo",
+            category="audio/partner/Sonilo",
             description="Generate music from a text prompt using Sonilo's AI model. "
             "Leave duration at 0 to let the model infer it from the prompt.",
             inputs=[

comfy_api_nodes/nodes_sora.py

@@ -34,7 +34,7 @@ class OpenAIVideoSora2(IO.ComfyNode):
         return IO.Schema(
             node_id="OpenAIVideoSora2",
             display_name="OpenAI Sora - Video (DEPRECATED)",
-            category="api node/video/Sora",
+            category="video/partner/Sora",
             description=(
                 "OpenAI video and audio generation.\n\n"
                 "DEPRECATION NOTICE: OpenAI will stop serving the Sora v2 API in September 2026. "

comfy_api_nodes/nodes_stability.py

@@ -62,7 +62,7 @@ class StabilityStableImageUltraNode(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityStableImageUltraNode",
             display_name="Stability AI Stable Image Ultra",
-            category="api node/image/Stability AI",
+            category="image/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.String.Input(
@@ -197,7 +197,7 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityStableImageSD_3_5Node",
             display_name="Stability AI Stable Diffusion 3.5 Image",
-            category="api node/image/Stability AI",
+            category="image/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.String.Input(
@@ -354,7 +354,7 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityUpscaleConservativeNode",
             display_name="Stability AI Upscale Conservative",
-            category="api node/image/Stability AI",
+            category="image/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("image"),
@@ -457,7 +457,7 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityUpscaleCreativeNode",
             display_name="Stability AI Upscale Creative",
-            category="api node/image/Stability AI",
+            category="image/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("image"),
@@ -578,7 +578,7 @@ class StabilityUpscaleFastNode(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityUpscaleFastNode",
             display_name="Stability AI Upscale Fast",
-            category="api node/image/Stability AI",
+            category="image/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Image.Input("image"),
@@ -630,7 +630,7 @@ class StabilityTextToAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityTextToAudio",
             display_name="Stability AI Text To Audio",
-            category="api node/audio/Stability AI",
+            category="audio/partner/Stability AI",
             essentials_category="Audio",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
@@ -708,7 +708,7 @@ class StabilityAudioToAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityAudioToAudio",
             display_name="Stability AI Audio To Audio",
-            category="api node/audio/Stability AI",
+            category="audio/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Combo.Input(
@@ -802,7 +802,7 @@ class StabilityAudioInpaint(IO.ComfyNode):
         return IO.Schema(
             node_id="StabilityAudioInpaint",
             display_name="Stability AI Audio Inpaint",
-            category="api node/audio/Stability AI",
+            category="audio/partner/Stability AI",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Combo.Input(

comfy_api_nodes/nodes_topaz.py

@@ -52,7 +52,7 @@ class TopazImageEnhance(IO.ComfyNode):
         return IO.Schema(
             node_id="TopazImageEnhance",
             display_name="Topaz Image Enhance",
-            category="api node/image/Topaz",
+            category="image/partner/Topaz",
             description="Industry-standard upscaling and image enhancement.",
             inputs=[
                 IO.Combo.Input("model", options=["Reimagine"]),
@@ -235,7 +235,7 @@ class TopazVideoEnhance(IO.ComfyNode):
         return IO.Schema(
             node_id="TopazVideoEnhance",
             display_name="Topaz Video Enhance (Legacy)",
-            category="api node/video/Topaz",
+            category="video/partner/Topaz",
             description="Breathe new life into video with powerful upscaling and recovery technology.",
             inputs=[
                 IO.Video.Input("video"),
@@ -475,7 +475,7 @@ class TopazVideoEnhanceV2(IO.ComfyNode):
         return IO.Schema(
             node_id="TopazVideoEnhanceV2",
             display_name="Topaz Video Enhance",
-            category="api node/video/Topaz",
+            category="video/partner/Topaz",
             description="Breathe new life into video with powerful upscaling and recovery technology.",
             inputs=[
                 IO.Video.Input("video"),

comfy_api_nodes/nodes_tripo.py

@@ -11,6 +11,9 @@ from comfy_api_nodes.apis.tripo import (
     TripoModelVersion,
     TripoMultiviewToModelRequest,
     TripoOrientation,
+    TripoP1ImageToModelRequest,
+    TripoP1MultiviewToModelRequest,
+    TripoP1TextToModelRequest,
     TripoRefineModelRequest,
     TripoStyle,
     TripoTaskResponse,
@@ -80,7 +83,7 @@ class TripoTextToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoTextToModelNode",
             display_name="Tripo: Text to Model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[
                 IO.String.Input("prompt", multiline=True),
                 IO.String.Input("negative_prompt", multiline=True, optional=True),
@@ -93,10 +96,22 @@ class TripoTextToModelNode(IO.ComfyNode):
                 IO.Int.Input("image_seed", default=42, optional=True, advanced=True),
                 IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
                 IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
+                IO.Combo.Input(
+                    "texture_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
                 IO.Int.Input("face_limit", default=-1, min=-1, max=2000000, optional=True, advanced=True),
                 IO.Boolean.Input("quad", default=False, optional=True, advanced=True),
-                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
+                IO.Combo.Input(
+                    "geometry_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -195,7 +210,7 @@ class TripoImageToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoImageToModelNode",
             display_name="Tripo: Image to Model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[
                 IO.Image.Input("image"),
                 IO.Combo.Input(
@@ -209,16 +224,36 @@ class TripoImageToModelNode(IO.ComfyNode):
                 IO.Boolean.Input("pbr", default=True, optional=True),
                 IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
                 IO.Combo.Input(
-                    "orientation", options=TripoOrientation, default=TripoOrientation.DEFAULT, optional=True, advanced=True
+                    "orientation",
+                    options=TripoOrientation,
+                    default=TripoOrientation.DEFAULT,
+                    optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
                 IO.Combo.Input(
-                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True
+                    "texture_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
+                IO.Combo.Input(
+                    "texture_alignment",
+                    default="original_image",
+                    options=["original_image", "geometry"],
+                    optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True, advanced=True),
                 IO.Boolean.Input("quad", default=False, optional=True, advanced=True),
-                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
+                IO.Combo.Input(
+                    "geometry_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -323,7 +358,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoMultiviewToModelNode",
             display_name="Tripo: Multiview to Model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[
                 IO.Image.Input("image"),
                 IO.Image.Input("image_left", optional=True),
@@ -346,13 +381,35 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
                 IO.Boolean.Input("pbr", default=True, optional=True),
                 IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
                 IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
                 IO.Combo.Input(
-                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True
+                    "texture_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
+                IO.Combo.Input(
+                    "texture_alignment",
+                    default="original_image",
+                    options=["original_image", "geometry"],
+                    optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True, advanced=True),
-                IO.Boolean.Input("quad", default=False, optional=True, advanced=True, tooltip="This parameter is deprecated and does nothing."),
-                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
+                IO.Boolean.Input(
+                    "quad",
+                    default=False,
+                    optional=True,
+                    advanced=True,
+                    tooltip="This parameter is deprecated and does nothing.",
+                ),
+                IO.Combo.Input(
+                    "geometry_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -461,15 +518,25 @@ class TripoTextureNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoTextureNode",
             display_name="Tripo: Texture model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[
                 IO.Custom("MODEL_TASK_ID").Input("model_task_id"),
                 IO.Boolean.Input("texture", default=True, optional=True),
                 IO.Boolean.Input("pbr", default=True, optional=True),
                 IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
                 IO.Combo.Input(
-                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True
+                    "texture_quality",
+                    default="standard",
+                    options=["standard", "detailed"],
+                    optional=True,
+                    advanced=True,
+                ),
+                IO.Combo.Input(
+                    "texture_alignment",
+                    default="original_image",
+                    options=["original_image", "geometry"],
+                    optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -528,7 +595,7 @@ class TripoRefineNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoRefineNode",
             display_name="Tripo: Refine Draft model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             description="Refine a draft model created by v1.4 Tripo models only.",
             inputs=[
                 IO.Custom("MODEL_TASK_ID").Input("model_task_id", tooltip="Must be a v1.4 Tripo model"),
@@ -568,7 +635,7 @@ class TripoRigNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoRigNode",
             display_name="Tripo: Rig model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[IO.Custom("MODEL_TASK_ID").Input("original_model_task_id")],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -605,7 +672,7 @@ class TripoRetargetNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoRetargetNode",
             display_name="Tripo: Retarget rigged model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[
                 IO.Custom("RIG_TASK_ID").Input("original_model_task_id"),
                 IO.Combo.Input(
@@ -626,7 +693,7 @@ class TripoRetargetNode(IO.ComfyNode):
                         "preset:hexapod:walk",
                         "preset:octopod:walk",
                         "preset:serpentine:march",
-                        "preset:aquatic:march"
+                        "preset:aquatic:march",
                     ],
                 ),
             ],
@@ -670,7 +737,7 @@ class TripoConversionNode(IO.ComfyNode):
         return IO.Schema(
             node_id="TripoConversionNode",
             display_name="Tripo: Convert model",
-            category="api node/3d/Tripo",
+            category="3d/partner/Tripo",
             inputs=[
                 IO.Custom("MODEL_TASK_ID,RIG_TASK_ID,RETARGET_TASK_ID").Input("original_model_task_id"),
                 IO.Combo.Input("format", options=["GLTF", "USDZ", "FBX", "OBJ", "STL", "3MF"]),
@@ -817,7 +884,7 @@ class TripoConversionNode(IO.ComfyNode):
         # Parse part_names from comma-separated string to list
         part_names_list = None
         if part_names and part_names.strip():
-            part_names_list = [name.strip() for name in part_names.split(',') if name.strip()]
+            part_names_list = [name.strip() for name in part_names.split(",") if name.strip()]
 
         response = await sync_op(
             cls,
@@ -848,6 +915,373 @@ class TripoConversionNode(IO.ComfyNode):
         return await poll_until_finished(cls, response, average_duration=30)
 
 
+def _p1_price_expr(*, geometry_credits: int, textured_credits: int, detailed_credits: int) -> str:
+    return (
+        "("
+        "  $mode := widgets.output_mode;"
+        '  $detailed := $lookup(widgets, "output_mode.texture_quality") = "detailed";'
+        f'  $credits := $mode = "geometry only" ? {geometry_credits} : ($detailed ? {detailed_credits} : {textured_credits});'
+        '  {"type":"usd","usd": $credits * 0.01, "format": {"approximate": true}}'
+        ")"
+    )
+
+
+def _p1_textured_inputs(*, include_image_alignment: bool) -> list:
+    """Inputs shown inside the 'Textured' branch of the P1 output_mode DynamicCombo."""
+    inputs: list = [
+        IO.Boolean.Input("pbr", default=True, tooltip="Include PBR maps. When on, base texture is forced on too."),
+        IO.Combo.Input("texture_quality", options=["standard", "detailed"], default="standard"),
+    ]
+    if include_image_alignment:
+        inputs.extend(
+            [
+                IO.Combo.Input(
+                    "texture_alignment",
+                    options=["original_image", "geometry"],
+                    default="original_image",
+                    tooltip="Prioritize visual fidelity to the source image, or alignment to the mesh geometry.",
+                ),
+                IO.Combo.Input(
+                    "orientation",
+                    options=["default", "align_image"],
+                    default="default",
+                    tooltip="Rotate the output to match the source image. Only applies when textured.",
+                ),
+            ]
+        )
+    inputs.append(IO.Int.Input("texture_seed", default=42, advanced=True))
+    return inputs
+
+
+def _build_p1_output_mode(*, include_image_alignment: bool) -> IO.DynamicCombo.Input:
+    return IO.DynamicCombo.Input(
+        "output_mode",
+        options=[
+            IO.DynamicCombo.Option("Geometry only", []),
+            IO.DynamicCombo.Option("Textured", _p1_textured_inputs(include_image_alignment=include_image_alignment)),
+        ],
+        tooltip='"Geometry only" returns an untextured mesh. "Textured" adds color/PBR maps.',
+    )
+
+
+def _resolve_p1_texture_fields(output_mode: dict) -> dict:
+    """Translate the output_mode DynamicCombo payload into P1 request fields.
+
+    pbr=true forces texture=true server-side, but we send both explicitly so the
+    intent is visible in the request body and logs.
+    """
+    mode = output_mode["output_mode"]
+    if mode == "Geometry only":
+        return {"texture": False, "pbr": False}
+    out = {
+        "texture": True,
+        "pbr": bool(output_mode.get("pbr", True)),
+        "texture_quality": output_mode.get("texture_quality", "standard"),
+        "texture_seed": output_mode.get("texture_seed"),
+    }
+    if "texture_alignment" in output_mode:
+        out["texture_alignment"] = output_mode["texture_alignment"]
+    if "orientation" in output_mode:
+        out["orientation"] = output_mode["orientation"]
+    return out
+
+
+def _p1_common_inputs() -> list:
+    """Inputs shared by all P1 nodes (placed after output_mode)."""
+    return [
+        IO.Int.Input(
+            "face_limit",
+            default=-1,
+            min=-1,
+            max=20000,
+            optional=True,
+            advanced=True,
+            tooltip="Target face count, 48-20000. -1 lets Tripo pick adaptively.",
+        ),
+        IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
+        IO.Boolean.Input(
+            "auto_size",
+            default=False,
+            optional=True,
+            advanced=True,
+            tooltip="Scale the output to approximate real-world meters.",
+        ),
+        IO.Boolean.Input(
+            "export_uv",
+            default=True,
+            optional=True,
+            advanced=True,
+            tooltip="UV unwrap during generation. Turn off for faster geometry-only runs.",
+        ),
+        IO.Boolean.Input(
+            "compress_geometry",
+            default=False,
+            optional=True,
+            advanced=True,
+            tooltip="Apply geometry-based compression. Decompress before editing.",
+        ),
+    ]
+
+
+def _build_p1_request_kwargs(
+    *,
+    output_mode: dict,
+    face_limit: int,
+    model_seed: int,
+    auto_size: bool,
+    export_uv: bool,
+    compress_geometry: bool,
+) -> dict:
+    """Common P1 request fields shared by all three node types."""
+    kwargs: dict = {
+        "model_seed": model_seed,
+        "face_limit": face_limit if face_limit != -1 else None,
+        "auto_size": auto_size,
+        "export_uv": export_uv,
+        "compress": "geometry" if compress_geometry else None,
+    }
+    kwargs.update(_resolve_p1_texture_fields(output_mode))
+    return kwargs
+
+
+class TripoP1TextToModelNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TripoP1TextToModelNode",
+            display_name="Tripo P1: Text to Model",
+            category="3d/partner/Tripo",
+            description="Tripo P1 text-to-3D. Optimized for low-poly, game-ready meshes with stable topology.",
+            inputs=[
+                IO.String.Input("prompt", multiline=True, tooltip="Up to 1024 characters."),
+                IO.String.Input("negative_prompt", multiline=True, optional=True, tooltip="Up to 255 characters."),
+                _build_p1_output_mode(include_image_alignment=False),
+                IO.Int.Input("image_seed", default=42, optional=True, advanced=True),
+                *_p1_common_inputs(),
+            ],
+            outputs=[
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["output_mode", "output_mode.texture_quality"]),
+                expr=_p1_price_expr(geometry_credits=30, textured_credits=40, detailed_credits=50),
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        output_mode: dict,
+        negative_prompt: str | None = None,
+        image_seed: int | None = None,
+        face_limit: int = -1,
+        model_seed: int | None = None,
+        auto_size: bool = False,
+        export_uv: bool = True,
+        compress_geometry: bool = False,
+    ) -> IO.NodeOutput:
+        if not prompt:
+            raise RuntimeError("Prompt is required")
+        common = _build_p1_request_kwargs(
+            output_mode=output_mode,
+            face_limit=face_limit,
+            model_seed=model_seed,
+            auto_size=auto_size,
+            export_uv=export_uv,
+            compress_geometry=compress_geometry,
+        )
+        request = TripoP1TextToModelRequest(
+            prompt=prompt,
+            negative_prompt=negative_prompt or None,
+            image_seed=image_seed,
+            **common,
+        )
+        response = await sync_op(
+            cls,
+            endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"),
+            response_model=TripoTaskResponse,
+            data=request,
+        )
+        return await poll_until_finished(cls, response, average_duration=60)
+
+
+class TripoP1ImageToModelNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TripoP1ImageToModelNode",
+            display_name="Tripo P1: Image to Model",
+            category="3d/partner/Tripo",
+            description="Tripo P1 image-to-3D. Optimized for low-poly, game-ready meshes.",
+            inputs=[
+                IO.Image.Input("image"),
+                _build_p1_output_mode(include_image_alignment=True),
+                IO.Boolean.Input(
+                    "enable_image_autofix",
+                    default=False,
+                    optional=True,
+                    advanced=True,
+                    tooltip="Pre-process the input image for better generation quality.",
+                ),
+                *_p1_common_inputs(),
+            ],
+            outputs=[
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["output_mode", "output_mode.texture_quality"]),
+                expr=_p1_price_expr(geometry_credits=40, textured_credits=50, detailed_credits=60),
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        output_mode: dict,
+        enable_image_autofix: bool = False,
+        face_limit: int = -1,
+        model_seed: int | None = None,
+        auto_size: bool = False,
+        export_uv: bool = True,
+        compress_geometry: bool = False,
+    ) -> IO.NodeOutput:
+        if image is None:
+            raise RuntimeError("Image is required")
+        tripo_file = TripoFileReference(
+            root=TripoUrlReference(
+                url=(await upload_images_to_comfyapi(cls, image, max_images=1))[0],
+                type="jpeg",
+            )
+        )
+        common = _build_p1_request_kwargs(
+            output_mode=output_mode,
+            face_limit=face_limit,
+            model_seed=model_seed,
+            auto_size=auto_size,
+            export_uv=export_uv,
+            compress_geometry=compress_geometry,
+        )
+        request = TripoP1ImageToModelRequest(
+            file=tripo_file,
+            enable_image_autofix=enable_image_autofix,
+            **common,
+        )
+        response = await sync_op(
+            cls,
+            endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"),
+            response_model=TripoTaskResponse,
+            data=request,
+        )
+        return await poll_until_finished(cls, response, average_duration=60)
+
+
+class TripoP1MultiviewToModelNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TripoP1MultiviewToModelNode",
+            display_name="Tripo P1: Multiview to Model",
+            category="3d/partner/Tripo",
+            description="Tripo P1 multiview-to-3D from 2-4 reference images in [front, left, back, right] order. "
+            "Front is required; any combination of the other three may be omitted.",
+            inputs=[
+                IO.Image.Input("image", tooltip="Front view (0°). Required."),
+                IO.Image.Input(
+                    "image_left",
+                    optional=True,
+                    tooltip="Left view (90°), i.e. the subject's left side.",
+                ),
+                IO.Image.Input("image_back", optional=True, tooltip="Back view (180°)."),
+                IO.Image.Input(
+                    "image_right",
+                    optional=True,
+                    tooltip="Right view (270°), i.e. the subject's right side.",
+                ),
+                _build_p1_output_mode(include_image_alignment=True),
+                *_p1_common_inputs(),
+            ],
+            outputs=[
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["output_mode", "output_mode.texture_quality"]),
+                expr=_p1_price_expr(geometry_credits=40, textured_credits=50, detailed_credits=60),
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        output_mode: dict,
+        image_left: Input.Image | None = None,
+        image_back: Input.Image | None = None,
+        image_right: Input.Image | None = None,
+        face_limit: int = -1,
+        model_seed: int | None = None,
+        auto_size: bool = False,
+        export_uv: bool = True,
+        compress_geometry: bool = False,
+    ) -> IO.NodeOutput:
+        views = [image, image_left, image_back, image_right]
+        if sum(1 for v in views if v is not None) < 2:
+            raise RuntimeError("Tripo P1 multiview requires at least 2 images (front plus one of left/back/right).")
+
+        files: list[TripoFileReference] = []
+        for view in views:
+            if view is None:
+                files.append(TripoFileReference(root=TripoFileEmptyReference()))
+                continue
+            url = (await upload_images_to_comfyapi(cls, view, max_images=1))[0]
+            files.append(TripoFileReference(root=TripoUrlReference(url=url, type="jpeg")))
+
+        common = _build_p1_request_kwargs(
+            output_mode=output_mode,
+            face_limit=face_limit,
+            model_seed=model_seed,
+            auto_size=auto_size,
+            export_uv=export_uv,
+            compress_geometry=compress_geometry,
+        )
+        request = TripoP1MultiviewToModelRequest(files=files, **common)
+        response = await sync_op(
+            cls,
+            endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"),
+            response_model=TripoTaskResponse,
+            data=request,
+        )
+        return await poll_until_finished(cls, response, average_duration=80)
+
+
 class TripoExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -855,6 +1289,9 @@ class TripoExtension(ComfyExtension):
             TripoTextToModelNode,
             TripoImageToModelNode,
             TripoMultiviewToModelNode,
+            TripoP1TextToModelNode,
+            TripoP1ImageToModelNode,
+            TripoP1MultiviewToModelNode,
             TripoTextureNode,
             TripoRefineNode,
             TripoRigNode,

comfy_api_nodes/nodes_veo2.py

@@ -45,7 +45,7 @@ class VeoVideoGenerationNode(IO.ComfyNode):
         return IO.Schema(
             node_id="VeoVideoGenerationNode",
             display_name="Google Veo 2 Video Generation",
-            category="api node/video/Veo",
+            category="video/partner/Veo",
             description="Generates videos from text prompts using Google's Veo 2 API",
             inputs=[
                 IO.String.Input(
@@ -256,7 +256,7 @@ class Veo3VideoGenerationNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Veo3VideoGenerationNode",
             display_name="Google Veo 3 Video Generation",
-            category="api node/video/Veo",
+            category="video/partner/Veo",
             description="Generates videos from text prompts using Google's Veo 3 API",
             inputs=[
                 IO.String.Input(
@@ -468,7 +468,7 @@ class Veo3FirstLastFrameNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Veo3FirstLastFrameNode",
             display_name="Google Veo 3 First-Last-Frame to Video",
-            category="api node/video/Veo",
+            category="video/partner/Veo",
             description="Generate video using prompt and first and last frames.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_vidu.py

@@ -71,7 +71,7 @@ class ViduTextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ViduTextToVideoNode",
             display_name="Vidu Text To Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate video from a text prompt",
             inputs=[
                 IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"),
@@ -169,7 +169,7 @@ class ViduImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ViduImageToVideoNode",
             display_name="Vidu Image To Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate video from image and optional prompt",
             inputs=[
                 IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"),
@@ -273,7 +273,7 @@ class ViduReferenceVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ViduReferenceVideoNode",
             display_name="Vidu Reference To Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate video from multiple images and a prompt",
             inputs=[
                 IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"),
@@ -388,7 +388,7 @@ class ViduStartEndToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ViduStartEndToVideoNode",
             display_name="Vidu Start End To Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video from start and end frames and a prompt",
             inputs=[
                 IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"),
@@ -492,7 +492,7 @@ class Vidu2TextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu2TextToVideoNode",
             display_name="Vidu2 Text-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate video from a text prompt",
             inputs=[
                 IO.Combo.Input("model", options=["viduq2"]),
@@ -584,7 +584,7 @@ class Vidu2ImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu2ImageToVideoNode",
             display_name="Vidu2 Image-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video from an image and an optional prompt.",
             inputs=[
                 IO.Combo.Input("model", options=["viduq2-pro-fast", "viduq2-pro", "viduq2-turbo"]),
@@ -714,7 +714,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu2ReferenceVideoNode",
             display_name="Vidu2 Reference-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video from multiple reference images and a prompt.",
             inputs=[
                 IO.Combo.Input("model", options=["viduq2"]),
@@ -849,7 +849,7 @@ class Vidu2StartEndToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu2StartEndToVideoNode",
             display_name="Vidu2 Start/End Frame-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video from a start frame, an end frame, and a prompt.",
             inputs=[
                 IO.Combo.Input("model", options=["viduq2-pro-fast", "viduq2-pro", "viduq2-turbo"]),
@@ -969,7 +969,7 @@ class ViduExtendVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ViduExtendVideoNode",
             display_name="Vidu Video Extension",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Extend an existing video by generating additional frames.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1138,7 +1138,7 @@ class ViduMultiFrameVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="ViduMultiFrameVideoNode",
             display_name="Vidu Multi-Frame Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video with multiple keyframe transitions.",
             inputs=[
                 IO.Combo.Input("model", options=["viduq2-pro", "viduq2-turbo"]),
@@ -1284,7 +1284,7 @@ class Vidu3TextToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu3TextToVideoNode",
             display_name="Vidu Q3 Text-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate video from a text prompt.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1429,7 +1429,7 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu3ImageToVideoNode",
             display_name="Vidu Q3 Image-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video from an image and an optional prompt.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1571,7 +1571,7 @@ class Vidu3StartEndToVideoNode(IO.ComfyNode):
         return IO.Schema(
             node_id="Vidu3StartEndToVideoNode",
             display_name="Vidu Q3 Start/End Frame-to-Video Generation",
-            category="api node/video/Vidu",
+            category="video/partner/Vidu",
             description="Generate a video from a start frame, an end frame, and a prompt.",
             inputs=[
                 IO.DynamicCombo.Input(

comfy_api_nodes/nodes_wan.py

@@ -61,7 +61,7 @@ class WanTextToImageApi(IO.ComfyNode):
         return IO.Schema(
             node_id="WanTextToImageApi",
             display_name="Wan Text to Image",
-            category="api node/image/Wan",
+            category="image/partner/Wan",
             description="Generates an image based on a text prompt.",
             inputs=[
                 IO.Combo.Input(
@@ -184,7 +184,7 @@ class WanImageToImageApi(IO.ComfyNode):
         return IO.Schema(
             node_id="WanImageToImageApi",
             display_name="Wan Image to Image",
-            category="api node/image/Wan",
+            category="image/partner/Wan",
             description="Generates an image from one or two input images and a text prompt. "
             "The output image is currently fixed at 1.6 MP, and its aspect ratio matches the input image(s).",
             inputs=[
@@ -312,7 +312,7 @@ class WanTextToVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="WanTextToVideoApi",
             display_name="Wan Text to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generates a video based on a text prompt.",
             inputs=[
                 IO.Combo.Input(
@@ -495,7 +495,7 @@ class WanImageToVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="WanImageToVideoApi",
             display_name="Wan Image to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generates a video from the first frame and a text prompt.",
             inputs=[
                 IO.Combo.Input(
@@ -674,7 +674,7 @@ class WanReferenceVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="WanReferenceVideoApi",
             display_name="Wan Reference to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Use the character and voice from input videos, combined with a prompt, "
             "to generate a new video that maintains character consistency.",
             inputs=[
@@ -828,7 +828,7 @@ class Wan2TextToVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="Wan2TextToVideoApi",
             display_name="Wan 2.7 Text to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generates a video based on a text prompt using the Wan 2.7 model.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -981,7 +981,7 @@ class Wan2ImageToVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="Wan2ImageToVideoApi",
             display_name="Wan 2.7 Image to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generate a video from a first-frame image, with optional last-frame image and audio.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1152,7 +1152,7 @@ class Wan2VideoContinuationApi(IO.ComfyNode):
         return IO.Schema(
             node_id="Wan2VideoContinuationApi",
             display_name="Wan 2.7 Video Continuation",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Continue a video from where it left off, with optional last-frame control.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1319,7 +1319,7 @@ class Wan2VideoEditApi(IO.ComfyNode):
         return IO.Schema(
             node_id="Wan2VideoEditApi",
             display_name="Wan 2.7 Video Edit",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Edit a video using text instructions, reference images, or style transfer.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1477,7 +1477,7 @@ class Wan2ReferenceVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="Wan2ReferenceVideoApi",
             display_name="Wan 2.7 Reference to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generate a video featuring a person or object from reference materials. "
             "Supports single-character performances and multi-character interactions.",
             inputs=[
@@ -1651,7 +1651,7 @@ class HappyHorseTextToVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="HappyHorseTextToVideoApi",
             display_name="HappyHorse Text to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generates a video based on a text prompt using the HappyHorse model.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1775,7 +1775,7 @@ class HappyHorseImageToVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="HappyHorseImageToVideoApi",
             display_name="HappyHorse Image to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generate a video from a first-frame image using the HappyHorse model.",
             inputs=[
                 IO.DynamicCombo.Input(
@@ -1905,7 +1905,7 @@ class HappyHorseVideoEditApi(IO.ComfyNode):
         return IO.Schema(
             node_id="HappyHorseVideoEditApi",
             display_name="HappyHorse Video Edit",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Edit a video using text instructions or reference images with the HappyHorse model. "
             "Output duration is 3-15s and matches the input video; inputs longer than 15s are truncated.",
             inputs=[
@@ -2046,7 +2046,7 @@ class HappyHorseReferenceVideoApi(IO.ComfyNode):
         return IO.Schema(
             node_id="HappyHorseReferenceVideoApi",
             display_name="HappyHorse Reference to Video",
-            category="api node/video/Wan",
+            category="video/partner/Wan",
             description="Generate a video featuring a person or object from reference materials with the HappyHorse "
             "model. Supports single-character performances and multi-character interactions.",
             inputs=[

comfy_api_nodes/nodes_wavespeed.py

@@ -27,7 +27,7 @@ class WavespeedFlashVSRNode(IO.ComfyNode):
         return IO.Schema(
             node_id="WavespeedFlashVSRNode",
             display_name="FlashVSR Video Upscale",
-            category="api node/video/WaveSpeed",
+            category="video/partner/WaveSpeed",
             description="Fast, high-quality video upscaler that "
             "boosts resolution and restores clarity for low-resolution or blurry footage.",
             inputs=[
@@ -98,7 +98,7 @@ class WavespeedImageUpscaleNode(IO.ComfyNode):
         return IO.Schema(
             node_id="WavespeedImageUpscaleNode",
             display_name="WaveSpeed Image Upscale",
-            category="api node/image/WaveSpeed",
+            category="image/partner/WaveSpeed",
             description="Boost image resolution and quality, upscaling photos to 4K or 8K for sharp, detailed results.",
             inputs=[
                 IO.Combo.Input("model", options=["SeedVR2", "Ultimate"]),

comfy_api_nodes/util/client.py

@@ -86,7 +86,7 @@ class _PollUIState:
 _RETRY_STATUS = {408, 500, 502, 503, 504}  # status 429 is handled separately
 COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"]
 FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"]
-QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing", "wait"]
+QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing", "wait", "in_queue"]
 
 
 async def sync_op(

comfy_extras/nodes_ace.py

@@ -11,7 +11,7 @@ class TextEncodeAceStepAudio(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TextEncodeAceStepAudio",
-            category="conditioning",
+            category="model/conditioning",
             inputs=[
                 IO.Clip.Input("clip"),
                 IO.String.Input("tags", multiline=True, dynamic_prompts=True),
@@ -33,7 +33,7 @@ class TextEncodeAceStepAudio15(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TextEncodeAceStepAudio1.5",
-            category="conditioning",
+            category="model/conditioning",
             inputs=[
                 IO.Clip.Input("clip"),
                 IO.String.Input("tags", multiline=True, dynamic_prompts=True),
@@ -67,7 +67,7 @@ class EmptyAceStepLatentAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="EmptyAceStepLatentAudio",
             display_name="Empty Ace Step 1.0 Latent Audio",
-            category="latent/audio",
+            category="model/latent/audio",
             inputs=[
                 IO.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
                 IO.Int.Input(
@@ -90,7 +90,7 @@ class EmptyAceStep15LatentAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="EmptyAceStep1.5LatentAudio",
             display_name="Empty Ace Step 1.5 Latent Audio",
-            category="latent/audio",
+            category="model/latent/audio",
             inputs=[
                 IO.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01),
                 IO.Int.Input(

comfy_extras/nodes_advanced_samplers.py

@@ -45,7 +45,7 @@ class SamplerLCMUpscale(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="SamplerLCMUpscale",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Float.Input("scale_ratio", default=1.0, min=0.1, max=20.0, step=0.01, advanced=True),
                 io.Int.Input("scale_steps", default=-1, min=-1, max=1000, step=1, advanced=True),
@@ -91,7 +91,7 @@ class SamplerLCM(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="SamplerLCM",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             description=("LCM sampler with tunable per-step noise. s_noise is a multiplier on the model's training noise scale"),
             inputs=[
                 io.Float.Input("s_noise", default=1.0, min=0.0, max=64.0, step=0.01,

comfy_extras/nodes_align_your_steps.py

@@ -29,7 +29,7 @@ class AlignYourStepsScheduler(io.ComfyNode):
         return io.Schema(
             node_id="AlignYourStepsScheduler",
             search_aliases=["AYS scheduler"],
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Combo.Input("model_type", options=["SD1", "SDXL", "SVD"]),
                 io.Int.Input("steps", default=10, min=1, max=10000),

comfy_extras/nodes_apg.py

@@ -16,7 +16,7 @@ class APG(io.ComfyNode):
         return io.Schema(
             node_id="APG",
             display_name="Adaptive Projected Guidance",
-            category="sampling/custom_sampling",
+            category="model/sampling/custom_sampling",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input(

comfy_extras/nodes_ar_video.py

@@ -19,7 +19,7 @@ class EmptyARVideoLatent(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyARVideoLatent",
-            category="latent/video",
+            category="model/latent/video",
             inputs=[
                 io.Int.Input("width", default=832, min=16, max=8192, step=16),
                 io.Int.Input("height", default=480, min=16, max=8192, step=16),
@@ -53,7 +53,7 @@ class SamplerARVideo(io.ComfyNode):
         return io.Schema(
             node_id="SamplerARVideo",
             display_name="Sampler AR Video",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Int.Input(
                     "num_frame_per_block",
@@ -85,7 +85,7 @@ class ARVideoI2V(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ARVideoI2V",
-            category="conditioning/video_models",
+            category="model/conditioning/video_models",
             inputs=[
                 io.Model.Input("model"),
                 io.Vae.Input("vae"),

comfy_extras/nodes_audio.py

@@ -16,7 +16,7 @@ class EmptyLatentAudio(IO.ComfyNode):
         return IO.Schema(
             node_id="EmptyLatentAudio",
             display_name="Empty Latent Audio",
-            category="latent/audio",
+            category="model/latent/audio",
             essentials_category="Audio",
             inputs=[
                 IO.Float.Input("seconds", default=47.6, min=1.0, max=1000.0, step=0.1),
@@ -41,7 +41,7 @@ class ConditioningStableAudio(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="ConditioningStableAudio",
-            category="conditioning",
+            category="model/conditioning",
             inputs=[
                 IO.Conditioning.Input("positive"),
                 IO.Conditioning.Input("negative"),
@@ -70,7 +70,7 @@ class VAEEncodeAudio(IO.ComfyNode):
             node_id="VAEEncodeAudio",
             search_aliases=["audio to latent"],
             display_name="VAE Encode Audio",
-            category="latent/audio",
+            category="model/latent/audio",
             inputs=[
                 IO.Audio.Input("audio"),
                 IO.Vae.Input("vae"),
@@ -115,7 +115,7 @@ class VAEDecodeAudio(IO.ComfyNode):
             node_id="VAEDecodeAudio",
             search_aliases=["latent to audio"],
             display_name="VAE Decode Audio",
-            category="latent/audio",
+            category="model/latent/audio",
             inputs=[
                 IO.Latent.Input("samples"),
                 IO.Vae.Input("vae"),
@@ -137,7 +137,7 @@ class VAEDecodeAudioTiled(IO.ComfyNode):
             node_id="VAEDecodeAudioTiled",
             search_aliases=["latent to audio"],
             display_name="VAE Decode Audio (Tiled)",
-            category="latent/audio",
+            category="model/latent/audio",
             inputs=[
                 IO.Latent.Input("samples"),
                 IO.Vae.Input("vae"),

comfy_extras/nodes_audio_encoder.py

@@ -11,7 +11,7 @@ class AudioEncoderLoader(io.ComfyNode):
         return io.Schema(
             node_id="AudioEncoderLoader",
             display_name="Load Audio Encoder",
-            category="loaders",
+            category="model/loaders",
             inputs=[
                 io.Combo.Input(
                     "audio_encoder_name",
@@ -36,7 +36,7 @@ class AudioEncoderEncode(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="AudioEncoderEncode",
-            category="conditioning",
+            category="model/conditioning",
             inputs=[
                 io.AudioEncoder.Input("audio_encoder"),
                 io.Audio.Input("audio"),

comfy_extras/nodes_bg_removal.py

@@ -11,7 +11,7 @@ class LoadBackgroundRemovalModel(IO.ComfyNode):
         return IO.Schema(
             node_id="LoadBackgroundRemovalModel",
             display_name="Load Background Removal Model",
-            category="loaders",
+            category="model/loaders",
             inputs=[
                 IO.Combo.Input("bg_removal_name", options=sorted(files), tooltip="The model used to remove backgrounds from images"),
             ],

comfy_extras/nodes_camera_trajectory.py

@@ -153,7 +153,7 @@ class WanCameraEmbedding(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="WanCameraEmbedding",
-            category="conditioning/video_models",
+            category="model/conditioning/video_models",
             inputs=[
                 io.Combo.Input(
                     "camera_pose",

comfy_extras/nodes_chroma_radiance.py

@@ -13,7 +13,7 @@ class EmptyChromaRadianceLatentImage(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="EmptyChromaRadianceLatentImage",
-            category="latent/chroma_radiance",
+            category="model/latent/chroma_radiance",
             inputs=[
                 io.Int.Input(id="width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16),
                 io.Int.Input(id="height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16),
@@ -33,7 +33,7 @@ class ChromaRadianceOptions(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="ChromaRadianceOptions",
-            category="model_patches/chroma_radiance",
+            category="model/patch/chroma_radiance",
             description="Allows setting advanced options for the Chroma Radiance model.",
             inputs=[
                 io.Model.Input(id="model"),

comfy_extras/nodes_color.py

@@ -8,7 +8,7 @@ class ColorToRGBInt(io.ComfyNode):
         return io.Schema(
             node_id="ColorToRGBInt",
             display_name="Color to RGB Int",
-            category="utils",
+            category="utilities",
             description="Convert a color to a RGB integer value.",
             inputs=[
                 io.Color.Input("color"),

comfy_extras/nodes_context_windows.py

@@ -9,7 +9,7 @@ class ContextWindowsManualNode(io.ComfyNode):
         return io.Schema(
             node_id="ContextWindowsManual",
             display_name="Context Windows (Manual)",
-            category="model_patches",
+            category="model/patch",
             description="Manually set context windows.",
             inputs=[
                 io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),

comfy_extras/nodes_controlnet.py

@@ -9,7 +9,7 @@ class SetUnionControlNetType(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SetUnionControlNetType",
-            category="conditioning/controlnet",
+            category="model/conditioning/controlnet",
             inputs=[
                 io.ControlNet.Input("control_net"),
                 io.Combo.Input("type", options=["auto"] + list(UNION_CONTROLNET_TYPES.keys())),
@@ -39,7 +39,7 @@ class ControlNetInpaintingAliMamaApply(io.ComfyNode):
         return io.Schema(
             node_id="ControlNetInpaintingAliMamaApply",
             search_aliases=["masked controlnet"],
-            category="conditioning/controlnet",
+            category="model/conditioning/controlnet",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Conditioning.Input("negative"),

comfy_extras/nodes_cosmos.py

@@ -13,7 +13,7 @@ class EmptyCosmosLatentVideo(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="EmptyCosmosLatentVideo",
-            category="latent/video",
+            category="model/latent/video",
             inputs=[
                 io.Int.Input("width", default=1280, min=16, max=nodes.MAX_RESOLUTION, step=16),
                 io.Int.Input("height", default=704, min=16, max=nodes.MAX_RESOLUTION, step=16),
@@ -45,7 +45,7 @@ class CosmosImageToVideoLatent(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="CosmosImageToVideoLatent",
-            category="conditioning/inpaint",
+            category="model/conditioning/inpaint",
             inputs=[
                 io.Vae.Input("vae"),
                 io.Int.Input("width", default=1280, min=16, max=nodes.MAX_RESOLUTION, step=16),
@@ -88,7 +88,7 @@ class CosmosPredict2ImageToVideoLatent(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="CosmosPredict2ImageToVideoLatent",
-            category="conditioning/inpaint",
+            category="model/conditioning/inpaint",
             inputs=[
                 io.Vae.Input("vae"),
                 io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16),

comfy_extras/nodes_curve.py

@@ -11,7 +11,7 @@ class CurveEditor(io.ComfyNode):
         return io.Schema(
             node_id="CurveEditor",
             display_name="Curve Editor",
-            category="utils",
+            category="utilities",
             inputs=[
                 io.Curve.Input("curve"),
                 io.Histogram.Input("histogram", optional=True),
@@ -38,7 +38,7 @@ class ImageHistogram(io.ComfyNode):
         return io.Schema(
             node_id="ImageHistogram",
             display_name="Image Histogram",
-            category="utils",
+            category="utilities",
             inputs=[
                 io.Image.Input("image"),
             ],

comfy_extras/nodes_custom_sampler.py

@@ -17,7 +17,7 @@ class BasicScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="BasicScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Model.Input("model"),
                 io.Combo.Input("scheduler", options=comfy.samplers.SCHEDULER_NAMES),
@@ -47,7 +47,7 @@ class KarrasScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="KarrasScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Int.Input("steps", default=20, min=1, max=10000),
                 io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True),
@@ -69,7 +69,7 @@ class ExponentialScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ExponentialScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Int.Input("steps", default=20, min=1, max=10000),
                 io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True),
@@ -90,7 +90,7 @@ class PolyexponentialScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PolyexponentialScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Int.Input("steps", default=20, min=1, max=10000),
                 io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True),
@@ -112,7 +112,7 @@ class LaplaceScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="LaplaceScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Int.Input("steps", default=20, min=1, max=10000),
                 io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True),
@@ -136,7 +136,7 @@ class SDTurboScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SDTurboScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Model.Input("model"),
                 io.Int.Input("steps", default=1, min=1, max=10),
@@ -160,7 +160,7 @@ class BetaSamplingScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="BetaSamplingScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Model.Input("model"),
                 io.Int.Input("steps", default=20, min=1, max=10000),
@@ -182,7 +182,7 @@ class VPScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="VPScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Int.Input("steps", default=20, min=1, max=10000),
                 io.Float.Input("beta_d", default=19.9, min=0.0, max=5000.0, step=0.01, round=False, advanced=True), #TODO: fix default values
@@ -204,7 +204,7 @@ class SplitSigmas(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SplitSigmas",
-            category="sampling/sigmas",
+            category="model/sampling/sigmas",
             inputs=[
                 io.Sigmas.Input("sigmas"),
                 io.Int.Input("step", default=0, min=0, max=10000),
@@ -228,7 +228,7 @@ class SplitSigmasDenoise(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SplitSigmasDenoise",
-            category="sampling/sigmas",
+            category="model/sampling/sigmas",
             inputs=[
                 io.Sigmas.Input("sigmas"),
                 io.Float.Input("denoise", default=1.0, min=0.0, max=1.0, step=0.01),
@@ -254,7 +254,7 @@ class FlipSigmas(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="FlipSigmas",
-            category="sampling/sigmas",
+            category="model/sampling/sigmas",
             inputs=[io.Sigmas.Input("sigmas")],
             outputs=[io.Sigmas.Output()]
         )
@@ -276,7 +276,7 @@ class SetFirstSigma(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SetFirstSigma",
-            category="sampling/sigmas",
+            category="model/sampling/sigmas",
             inputs=[
                 io.Sigmas.Input("sigmas"),
                 io.Float.Input("sigma", default=136.0, min=0.0, max=20000.0, step=0.001, round=False),
@@ -298,7 +298,7 @@ class ExtendIntermediateSigmas(io.ComfyNode):
         return io.Schema(
             node_id="ExtendIntermediateSigmas",
             search_aliases=["interpolate sigmas"],
-            category="sampling/sigmas",
+            category="model/sampling/sigmas",
             inputs=[
                 io.Sigmas.Input("sigmas"),
                 io.Int.Input("steps", default=2, min=1, max=100),
@@ -351,7 +351,7 @@ class SamplingPercentToSigma(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplingPercentToSigma",
-            category="sampling/sigmas",
+            category="model/sampling/sigmas",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input("sampling_percent", default=0.0, min=0.0, max=1.0, step=0.0001),
@@ -379,7 +379,7 @@ class KSamplerSelect(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="KSamplerSelect",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[io.Combo.Input("sampler_name", options=comfy.samplers.SAMPLER_NAMES)],
             outputs=[io.Sampler.Output()]
         )
@@ -396,7 +396,7 @@ class SamplerDPMPP_3M_SDE(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerDPMPP_3M_SDE",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
                 io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
@@ -421,7 +421,7 @@ class SamplerDPMPP_2M_SDE(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerDPMPP_2M_SDE",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Combo.Input("solver_type", options=['midpoint', 'heun']),
                 io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
@@ -448,7 +448,7 @@ class SamplerDPMPP_SDE(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerDPMPP_SDE",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
                 io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
@@ -474,7 +474,7 @@ class SamplerDPMPP_2S_Ancestral(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerDPMPP_2S_Ancestral",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False),
                 io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False),
@@ -494,7 +494,7 @@ class SamplerEulerAncestral(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerEulerAncestral",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
                 io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
@@ -515,7 +515,7 @@ class SamplerEulerAncestralCFGPP(io.ComfyNode):
         return io.Schema(
             node_id="SamplerEulerAncestralCFGPP",
             display_name="SamplerEulerAncestralCFG++",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Float.Input("eta", default=1.0, min=0.0, max=1.0, step=0.01, round=False),
                 io.Float.Input("s_noise", default=1.0, min=0.0, max=10.0, step=0.01, round=False),
@@ -537,7 +537,7 @@ class SamplerLMS(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerLMS",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[io.Int.Input("order", default=4, min=1, max=100, advanced=True)],
             outputs=[io.Sampler.Output()]
         )
@@ -554,7 +554,7 @@ class SamplerDPMAdaptative(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerDPMAdaptative",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Int.Input("order", default=3, min=2, max=3, advanced=True),
                 io.Float.Input("rtol", default=0.05, min=0.0, max=100.0, step=0.01, round=False, advanced=True),
@@ -585,7 +585,7 @@ class SamplerER_SDE(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerER_SDE",
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Combo.Input("solver_type", options=["ER-SDE", "Reverse-time SDE", "ODE"]),
                 io.Int.Input("max_stage", default=3, min=1, max=3, advanced=True),
@@ -623,7 +623,7 @@ class SamplerSASolver(io.ComfyNode):
         return io.Schema(
             node_id="SamplerSASolver",
             search_aliases=["sde"],
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input("eta", default=1.0, min=0.0, max=10.0, step=0.01, round=False, advanced=True),
@@ -668,7 +668,7 @@ class SamplerSEEDS2(io.ComfyNode):
         return io.Schema(
             node_id="SamplerSEEDS2",
             search_aliases=["sde", "exp heun"],
-            category="sampling/samplers",
+            category="model/sampling/samplers",
             inputs=[
                 io.Combo.Input("solver_type", options=["phi_1", "phi_2"]),
                 io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="Stochastic strength", advanced=True),
@@ -727,7 +727,7 @@ class SamplerCustom(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerCustom",
-            category="sampling/custom_sampling",
+            category="model/sampling/custom_sampling",
             inputs=[
                 io.Model.Input("model"),
                 io.Boolean.Input("add_noise", default=True, advanced=True),
@@ -795,7 +795,7 @@ class BasicGuider(io.ComfyNode):
         return io.Schema(
             node_id="BasicGuider",
             display_name="Basic Guider",
-            category="sampling/guiders",
+            category="model/sampling/guiders",
             inputs=[
                 io.Model.Input("model"),
                 io.Conditioning.Input("conditioning"),
@@ -817,7 +817,7 @@ class CFGGuider(io.ComfyNode):
         return io.Schema(
             node_id="CFGGuider",
             display_name="CFG Guider",
-            category="sampling/guiders",
+            category="model/sampling/guiders",
             inputs=[
                 io.Model.Input("model"),
                 io.Conditioning.Input("positive"),
@@ -872,7 +872,7 @@ class DualCFGGuider(io.ComfyNode):
             node_id="DualCFGGuider",
             search_aliases=["dual prompt guidance"],
             display_name="Dual CFG Guider",
-            category="sampling/guiders",
+            category="model/sampling/guiders",
             inputs=[
                 io.Model.Input("model"),
                 io.Conditioning.Input("cond1"),
@@ -900,7 +900,7 @@ class DisableNoise(io.ComfyNode):
         return io.Schema(
             node_id="DisableNoise",
             search_aliases=["zero noise"],
-            category="sampling/noise",
+            category="model/sampling/noise",
             inputs=[],
             outputs=[io.Noise.Output()]
         )
@@ -917,7 +917,7 @@ class RandomNoise(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RandomNoise",
-            category="sampling/noise",
+            category="model/sampling/noise",
             inputs=[io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True)],
             outputs=[io.Noise.Output()]
         )
@@ -934,7 +934,7 @@ class SamplerCustomAdvanced(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerCustomAdvanced",
-            category="sampling/custom_sampling",
+            category="model/sampling/custom_sampling",
             inputs=[
                 io.Noise.Input("noise"),
                 io.Guider.Input("guider"),

comfy_extras/nodes_dataset.py

@@ -157,7 +157,7 @@ class LoadImageTextDataSetFromFolderNode(io.ComfyNode):
         return io.NodeOutput(output_tensor, captions)
 
 
-def save_images_to_folder(image_list, output_dir, prefix="image"):
+def save_images_to_folder(image_list, output_dir, prefix="image", overwrite=True):
     """Utility function to save a list of image tensors to disk.
 
     Args:
@@ -197,7 +197,11 @@ def save_images_to_folder(image_list, output_dir, prefix="image"):
             raise ValueError(f"Expected torch.Tensor, got {type(img_tensor)}")
 
         # Save image
-        filename = f"{prefix}_{idx:05d}.png"
+        if overwrite:
+            filename = f"{prefix}_{idx:05d}.png"
+        else:
+            _, _, counter, _, resolved_prefix = folder_paths.get_save_image_path(prefix, output_dir)
+            filename = f"{resolved_prefix}_{counter:05}_{idx:05d}.png"
         filepath = os.path.join(output_dir, filename)
         img.save(filepath)
         saved_files.append(filename)
@@ -230,19 +234,26 @@ class SaveImageDataSetToFolderNode(io.ComfyNode):
                     tooltip="Prefix for saved image filenames.",
                     advanced=True,
                 ),
+                io.Combo.Input(
+                    "mode",
+                    default="overwrite",
+                    options=["overwrite", "increment"],
+                    tooltip="Whether to overwrite existing files or increment filenames to avoid overwriting."
+                ),
             ],
             outputs=[],
             is_deprecated=True,  # This node is redundant and superseded by existing Save Image nodes where the target folder can be specified in the filename_prefix
         )
 
     @classmethod
-    def execute(cls, images, folder_name, filename_prefix):
+    def execute(cls, images, folder_name, filename_prefix, mode):
         # Extract scalar values
         folder_name = folder_name[0]
         filename_prefix = filename_prefix[0]
+        mode = mode[0]
 
         output_dir = os.path.join(folder_paths.get_output_directory(), folder_name)
-        saved_files = save_images_to_folder(images, output_dir, filename_prefix)
+        saved_files = save_images_to_folder(images, output_dir, filename_prefix, mode=='overwrite')
 
         logging.info(f"Saved {len(saved_files)} images to {output_dir}.")
         return io.NodeOutput()
@@ -278,18 +289,25 @@ class SaveImageTextDataSetToFolderNode(io.ComfyNode):
                     tooltip="Prefix for saved image filenames.",
                     advanced=True,
                 ),
+                io.Combo.Input(
+                    "mode",
+                    default="overwrite",
+                    options=["overwrite", "increment"],
+                    tooltip="Whether to overwrite existing files or increment filenames to avoid overwriting."
+                ),
             ],
             outputs=[],
         )
 
     @classmethod
-    def execute(cls, images, folder_name, filename_prefix, texts=None):
+    def execute(cls, images, folder_name, filename_prefix, mode, texts=None):
         # Extract scalar values
         folder_name = folder_name[0]
         filename_prefix = filename_prefix[0]
+        mode = mode[0]
 
         output_dir = os.path.join(folder_paths.get_output_directory(), folder_name)
-        saved_files = save_images_to_folder(images, output_dir, filename_prefix)
+        saved_files = save_images_to_folder(images, output_dir, filename_prefix, mode=='overwrite')
 
         # Save captions
         if texts:
@@ -574,7 +592,7 @@ class TextProcessingNode(io.ComfyNode):
         return io.Schema(
             node_id=cls.node_id,
             display_name=cls.display_name or cls.node_id,
-            category="dataset/text",
+            category="text",
             is_experimental=True,
             is_input_list=is_group,  # True for group, False for individual
             inputs=inputs,
@@ -1208,7 +1226,7 @@ class ResolutionBucket(io.ComfyNode):
             node_id="ResolutionBucket",
             search_aliases=["bucket by resolution", "group by resolution", "batch by resolution"],
             display_name="Resolution Bucket",
-            category="training",
+            category="model/training",
             description="Group latents and conditionings into buckets",
             is_experimental=True,
             is_input_list=True,
@@ -1302,7 +1320,7 @@ class MakeTrainingDataset(io.ComfyNode):
             node_id="MakeTrainingDataset",
             search_aliases=["encode dataset"],
             display_name="Make Training Dataset",
-            category="training",
+            category="model/training",
             description="Encode images with VAE and texts with CLIP to create a training dataset of latents and conditionings.",
             is_experimental=True,
             is_input_list=True,  # images and texts as lists
@@ -1390,7 +1408,7 @@ class SaveTrainingDataset(io.ComfyNode):
             node_id="SaveTrainingDataset",
             search_aliases=["export dataset", "save dataset"],
             display_name="Save Training Dataset",
-            category="training",
+            category="model/training",
             description="Save encoded training dataset (latents + conditioning) to disk for efficient loading during training.",
             is_experimental=True,
             is_output_node=True,
@@ -1493,7 +1511,7 @@ class LoadTrainingDataset(io.ComfyNode):
             node_id="LoadTrainingDataset",
             search_aliases=["import dataset", "training data"],
             display_name="Load Training Dataset",
-            category="training",
+            category="model/training",
             description="Load encoded training dataset (latents + conditioning) from disk for use in training.",
             is_experimental=True,
             inputs=[

comfy_extras/nodes_eps.py

@@ -18,7 +18,7 @@ class EpsilonScaling(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="Epsilon Scaling",
-            category="model_patches/unet",
+            category="model/patch/unet",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input(
@@ -84,7 +84,7 @@ class TemporalScoreRescaling(io.ComfyNode):
         return io.Schema(
             node_id="TemporalScoreRescaling",
             display_name="TSR - Temporal Score Rescaling",
-            category="model_patches/unet",
+            category="model/patch/unet",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input(

comfy_extras/nodes_flux.py

@@ -40,7 +40,7 @@ class EmptyFlux2LatentImage(io.ComfyNode):
         return io.Schema(
             node_id="EmptyFlux2LatentImage",
             display_name="Empty Flux 2 Latent",
-            category="latent",
+            category="model/latent",
             inputs=[
                 io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16),
                 io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16),
@@ -215,7 +215,7 @@ class Flux2Scheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="Flux2Scheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Int.Input("steps", default=20, min=1, max=4096),
                 io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1),

comfy_extras/nodes_frame_interpolation.py

@@ -19,7 +19,7 @@ class FrameInterpolationModelLoader(io.ComfyNode):
         return io.Schema(
             node_id="FrameInterpolationModelLoader",
             display_name="Load Frame Interpolation Model",
-            category="loaders",
+            category="model/loaders",
             inputs=[
                 io.Combo.Input("model_name", options=folder_paths.get_filename_list("frame_interpolation"),
                                tooltip="Select a frame interpolation model to load. Models must be placed in the 'frame_interpolation' folder."),

comfy_extras/nodes_freelunch.py

@@ -29,7 +29,7 @@ class FreeU(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="FreeU",
-            category="model_patches/unet",
+            category="model/patch/unet",
             inputs=[
                 IO.Model.Input("model"),
                 IO.Float.Input("b1", default=1.1, min=0.0, max=10.0, step=0.01, advanced=True),
@@ -76,7 +76,7 @@ class FreeU_V2(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="FreeU_V2",
-            category="model_patches/unet",
+            category="model/patch/unet",
             inputs=[
                 IO.Model.Input("model"),
                 IO.Float.Input("b1", default=1.3, min=0.0, max=10.0, step=0.01, advanced=True),

comfy_extras/nodes_gits.py

@@ -340,7 +340,7 @@ class GITSScheduler(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="GITSScheduler",
-            category="sampling/schedulers",
+            category="model/sampling/schedulers",
             inputs=[
                 io.Float.Input("coeff", default=1.20, min=0.80, max=1.50, step=0.05, advanced=True),
                 io.Int.Input("steps", default=10, min=2, max=1000),

comfy_extras/nodes_hidream_o1.py

@@ -14,7 +14,7 @@ class EmptyHiDreamO1LatentImage(io.ComfyNode):
         return io.Schema(
             node_id="EmptyHiDreamO1LatentImage",
             display_name="Empty HiDream-O1 Latent Image",
-            category="latent/image",
+            category="model/latent/image",
             description=(
                 "Empty pixel-space latent for HiDream-O1-Image. The model was "
                 "trained at ~4 megapixels; lower resolutions go off-distribution "
@@ -47,7 +47,7 @@ class HiDreamO1ReferenceImages(io.ComfyNode):
         return io.Schema(
             node_id="HiDreamO1ReferenceImages",
             display_name="HiDream-O1 Reference Images",
-            category="conditioning/image",
+            category="model/conditioning/image",
             description=(
                 "Attach 1-10 reference images to conditioning, one for edit instruction"
                 "or multiple for subject-driven personalization."

comfy_extras/nodes_hunyuan.py

@@ -41,7 +41,7 @@ class EmptyHunyuanLatentVideo(io.ComfyNode):
         return io.Schema(
             node_id="EmptyHunyuanLatentVideo",
             display_name="Empty HunyuanVideo 1.0 Latent",
-            category="latent/video",
+            category="model/latent/video",
             inputs=[
                 io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16),
                 io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
@@ -81,7 +81,7 @@ class HunyuanVideo15ImageToVideo(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="HunyuanVideo15ImageToVideo",
-            category="conditioning/video_models",
+            category="model/conditioning/video_models",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Conditioning.Input("negative"),
@@ -132,7 +132,7 @@ class HunyuanVideo15SuperResolution(io.ComfyNode):
         return io.Schema(
             node_id="HunyuanVideo15SuperResolution",
             display_name="Hunyuan Video 1.5 Super Resolution",
-            category="conditioning/video_models",
+            category="model/conditioning/video_models",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Conditioning.Input("negative"),
@@ -178,7 +178,7 @@ class LatentUpscaleModelLoader(io.ComfyNode):
         return io.Schema(
             node_id="LatentUpscaleModelLoader",
             display_name="Load Latent Upscale Model",
-            category="loaders",
+            category="model/loaders",
             inputs=[
                 io.Combo.Input("model_name", options=folder_paths.get_filename_list("latent_upscale_models")),
             ],
@@ -227,7 +227,7 @@ class HunyuanVideo15LatentUpscaleWithModel(io.ComfyNode):
         return io.Schema(
             node_id="HunyuanVideo15LatentUpscaleWithModel",
             display_name="Hunyuan Video 15 Latent Upscale With Model",
-            category="latent",
+            category="model/latent",
             inputs=[
                 io.LatentUpscaleModel.Input("model"),
                 io.Latent.Input("samples"),
@@ -308,7 +308,7 @@ class HunyuanImageToVideo(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="HunyuanImageToVideo",
-            category="conditioning/video_models",
+            category="model/conditioning/video_models",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Vae.Input("vae"),
@@ -359,7 +359,7 @@ class EmptyHunyuanImageLatent(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyHunyuanImageLatent",
-            category="latent",
+            category="model/latent",
             inputs=[
                 io.Int.Input("width", default=2048, min=64, max=nodes.MAX_RESOLUTION, step=32),
                 io.Int.Input("height", default=2048, min=64, max=nodes.MAX_RESOLUTION, step=32),
@@ -384,7 +384,7 @@ class HunyuanRefinerLatent(io.ComfyNode):
         return io.Schema(
             node_id="HunyuanRefinerLatent",
             display_name="Hunyuan Latent Refiner",
-            category="conditioning/video_models",
+            category="model/conditioning/video_models",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Conditioning.Input("negative"),

comfy_extras/nodes_hunyuan3d.py

@@ -12,7 +12,7 @@ class EmptyLatentHunyuan3Dv2(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="EmptyLatentHunyuan3Dv2",
-            category="latent/3d",
+            category="model/latent/3d",
             inputs=[
                 IO.Int.Input("resolution", default=3072, min=1, max=8192),
                 IO.Int.Input("batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."),
@@ -35,7 +35,7 @@ class Hunyuan3Dv2Conditioning(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="Hunyuan3Dv2Conditioning",
-            category="conditioning/3d_models",
+            category="model/conditioning/3d_models",
             inputs=[
                 IO.ClipVisionOutput.Input("clip_vision_output"),
             ],
@@ -60,7 +60,7 @@ class Hunyuan3Dv2ConditioningMultiView(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="Hunyuan3Dv2ConditioningMultiView",
-            category="conditioning/3d_models",
+            category="model/conditioning/3d_models",
             inputs=[
                 IO.ClipVisionOutput.Input("front", optional=True),
                 IO.ClipVisionOutput.Input("left", optional=True),
@@ -97,7 +97,7 @@ class VAEDecodeHunyuan3D(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="VAEDecodeHunyuan3D",
-            category="latent/3d",
+            category="model/latent/3d",
             inputs=[
                 IO.Latent.Input("samples"),
                 IO.Vae.Input("vae"),