Merge branch 'master' into fix/core/video-transcode-oom

.github/workflows/cla.yml

@@ -32,9 +32,11 @@ jobs:
           PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
           PR_AUTHOR: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
           BASE_ALLOWLIST: action@github.com,actions-user,ampagent,claude,comfy-pr-bot,GitHub Action,github-actions,github-actions[bot],Glary Bot,Glary-Bot,*[bot]
+        # For each commit emit the GitHub login when the author/committer email resolves to a GitHub account
+        # otherwise fall back to the raw git name.
         run: |
           others=$(gh api "repos/${{ github.repository }}/pulls/${PR_NUMBER}/commits" --paginate \
-            --jq '.[] | (.author.login // empty), (.committer.login // empty)' \
+            --jq '.[] | (.author.login // .commit.author.name // empty), (.committer.login // .commit.committer.name // empty)' \
             | sort -u | grep -vix "${PR_AUTHOR}" | paste -sd, -)
           if [ -n "$others" ]; then
             echo "allowlist=${BASE_ALLOWLIST},${others}" >> "$GITHUB_OUTPUT"
@@ -43,7 +45,7 @@ jobs:
           fi
 
       - name: CLA Assistant
-        # Run on PR events, on "recheck" comment, or when someone posts the exact signing phrase.
+        # Run on PR events, on "recheck" comment, or when someone posts the signing phrase.
         # IMPORTANT: this phrase must match `custom-pr-sign-comment` below.
         if: >
           github.event_name == 'pull_request_target' ||

app/model_manager.py

@@ -35,7 +35,11 @@ class ModelFileManager:
             for folder in model_types:
                 if folder in folder_black_list:
                     continue
-                output_folders.append({"name": folder, "folders": folder_paths.get_folder_paths(folder)})
+                output_folders.append({
+                    "name": folder,
+                    "folders": folder_paths.get_folder_paths(folder),
+                    "extensions": sorted(folder_paths.folder_names_and_paths[folder][1]),
+                })
             return web.json_response(output_folders)
 
         # NOTE: This is an experiment to replace `/models/{folder}`

comfy/cli_args.py

@@ -92,6 +92,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
 parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.")
+parser.add_argument("--disable-triton-backend", action="store_true", help="Force-disable the comfy-kitchen Triton backend, overriding the automatic ROCm/AMD default and --enable-triton-backend.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/latent_formats.py

@@ -779,6 +779,10 @@ class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
 
+class SeedVR2(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
 class ACEAudio15(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1

comfy/ldm/hidream_o1/attention.py

@@ -15,24 +15,24 @@ def make_two_pass_attention(ar_len: int, transformer_options=None):
     The AR pass goes through SDPA directand bypasses wrappers, it is only ~1% of T at typical edit sizes.
     """
 
-    def two_pass_attention(q, k, v, heads, **kwargs):
+    def two_pass_attention(q, k, v, heads, enable_gqa=False, **kwargs):
         B, H, T, D = q.shape
 
         if T < k.shape[2]: # KV-cache hot path: Q is shorter than K/V (cached AR prefix is in K/V only), all fresh Q positions are in the gen region, single full-attention call
-            out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options)
+            out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options, enable_gqa=enable_gqa)
         elif ar_len >= T:
-            out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=True)
+            out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=True, enable_gqa=enable_gqa)
         elif ar_len <= 0:
-            out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options)
+            out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options, enable_gqa=enable_gqa)
         else:
             out_ar = comfy.ops.scaled_dot_product_attention(
                 q[:, :, :ar_len], k[:, :, :ar_len], v[:, :, :ar_len],
-                attn_mask=None, dropout_p=0.0, is_causal=True,
+                attn_mask=None, dropout_p=0.0, is_causal=True, enable_gqa=enable_gqa,
             )
             out_gen = optimized_attention(
                 q[:, :, ar_len:], k, v, heads,
                 mask=None, skip_reshape=True, skip_output_reshape=True,
-                transformer_options=transformer_options,
+                transformer_options=transformer_options, enable_gqa=enable_gqa,
             )
             out = torch.cat([out_ar, out_gen], dim=2)
 

comfy/ldm/modules/attention.py

@@ -709,7 +709,7 @@ def attention3_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
     return out
 
 try:
-    @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
+    @torch.library.custom_op("comfy::flash_attn", mutates_args=())
     def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
                     dropout_p: float = 0.0, causal: bool = False, softmax_scale: float = -1.0) -> torch.Tensor:
         softmax_scale_arg = None if softmax_scale == -1.0 else softmax_scale

comfy/ldm/modules/diffusionmodules/model.py

@@ -22,7 +22,7 @@ def torch_cat_if_needed(xl, dim):
     else:
         return None
 
-def get_timestep_embedding(timesteps, embedding_dim):
+def get_timestep_embedding(timesteps, embedding_dim, flip_sin_to_cos=False, downscale_freq_shift=1):
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models:
     From Fairseq.
@@ -33,11 +33,13 @@ def get_timestep_embedding(timesteps, embedding_dim):
     assert len(timesteps.shape) == 1
 
     half_dim = embedding_dim // 2
-    emb = math.log(10000) / (half_dim - 1)
+    emb = math.log(10000) / (half_dim - downscale_freq_shift)
     emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
     emb = emb.to(device=timesteps.device)
     emb = timesteps.float()[:, None] * emb[None, :]
     emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
     if embedding_dim % 2 == 1:  # zero pad
         emb = torch.nn.functional.pad(emb, (0,1,0,0))
     return emb

comfy/ldm/pixeldit/model.py

@@ -197,6 +197,9 @@ class PixDiT_T2I(nn.Module):
         """Hook for subclasses to inject per-block state into the patch stream (e.g. PiD's LQ gate)."""
         return s
 
+    def _pre_pixel_blocks(self, s, **kwargs):
+        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))
@@ -226,6 +229,7 @@ class PixDiT_T2I(nn.Module):
             s, y_emb = blk(s, y_emb, condition, pos_img, pos_txt, None, transformer_options=transformer_options)
         s = F.silu(t_emb + s)
 
+        s = self._pre_pixel_blocks(s, **kwargs)
         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:

comfy/ldm/pixeldit/pid.py

@@ -13,15 +13,15 @@ from .model import PixDiT_T2I
 from .modules import precompute_freqs_cis_2d
 
 
-class SigmaAwareGatePerTokenPerDim(nn.Module):
+class SigmaAwareGate(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):
+    def __init__(self, dim: int, per_token: bool = False, dtype=None, device=None, operations=None):
         super().__init__()
-        self.content_proj = operations.Linear(dim * 2, dim, dtype=dtype, device=device)
+        self.content_proj = operations.Linear(dim * 2, 1 if per_token else 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:
@@ -36,15 +36,15 @@ class SigmaAwareGatePerTokenPerDim(nn.Module):
 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):
+    def __init__(self, channels: int, num_groups: int = 4, conv_padding_mode: str = "zeros", 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.Conv2d(channels, channels, kernel_size=3, padding=1, padding_mode=conv_padding_mode, 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),
+            operations.Conv2d(channels, channels, kernel_size=3, padding=1, padding_mode=conv_padding_mode, dtype=dtype, device=device),
         )
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -62,9 +62,13 @@ class LQProjection2D(nn.Module):
         patch_size: int = 16,
         sr_scale: int = 4,
         latent_spatial_down_factor: int = 8,
+        latent_unpatchify_factor: int = 1,
         num_res_blocks: int = 4,
         num_outputs: int = 7,
         interval: int = 2,
+        conv_padding_mode: str = "zeros",
+        gate_per_token: bool = False,
+        pit_output: bool = False,
         dtype=None, device=None, operations=None,
     ):
         super().__init__()
@@ -74,34 +78,38 @@ class LQProjection2D(nn.Module):
         self.patch_size = patch_size
         self.sr_scale = sr_scale
         self.latent_spatial_down_factor = latent_spatial_down_factor
+        self.latent_unpatchify_factor = latent_unpatchify_factor
         self.num_outputs = num_outputs
         self.interval = interval
 
-        z_to_patch_ratio = (sr_scale * latent_spatial_down_factor) / patch_size
+        effective_latent_channels = latent_channels // (latent_unpatchify_factor * latent_unpatchify_factor)
+        effective_spatial_down_factor = latent_spatial_down_factor // latent_unpatchify_factor
+        z_to_patch_ratio = (sr_scale * effective_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
+            latent_proj_in_ch = effective_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
+            latent_proj_in_ch = effective_latent_channels * fold_factor * fold_factor
 
         layers = [
-            operations.Conv2d(latent_proj_in_ch, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device),
+            operations.Conv2d(latent_proj_in_ch, hidden_dim, kernel_size=3, padding=1, padding_mode=conv_padding_mode, dtype=dtype, device=device),
             nn.SiLU(),
-            operations.Conv2d(hidden_dim, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device),
+            operations.Conv2d(hidden_dim, hidden_dim, kernel_size=3, padding=1, padding_mode=conv_padding_mode, dtype=dtype, device=device),
         ]
         for _ in range(num_res_blocks):
-            layers.append(ResBlock(hidden_dim, dtype=dtype, device=device, operations=operations))
+            layers.append(ResBlock(hidden_dim, conv_padding_mode=conv_padding_mode, 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.pit_head = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) if pit_output else None
         self.gate_modules = nn.ModuleList(
-            [SigmaAwareGatePerTokenPerDim(out_dim, dtype=dtype, device=device, operations=operations)
+            [SigmaAwareGate(out_dim, per_token=gate_per_token, dtype=dtype, device=device, operations=operations)
              for _ in range(num_outputs)]
         )
 
@@ -115,6 +123,11 @@ class LQProjection2D(nn.Module):
         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:
+        f = self.latent_unpatchify_factor
+        if f > 1:
+            B, C, H, W = lq_latent.shape
+            lq_latent = lq_latent.reshape(B, C // (f * f), f, f, H, W)
+            lq_latent = lq_latent.permute(0, 1, 4, 2, 5, 3).reshape(B, C // (f * f), H * f, W * f)
         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:
@@ -134,7 +147,10 @@ class LQProjection2D(nn.Module):
         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]
+        outputs = [head(tokens) for head in self.output_heads]
+        if self.pit_head is not None:
+            outputs.append(self.pit_head(tokens))
+        return outputs
 
 
 class PidNet(PixDiT_T2I):
@@ -148,6 +164,10 @@ class PidNet(PixDiT_T2I):
         lq_interval: int = 2,
         sr_scale: int = 4,
         latent_spatial_down_factor: int = 8,
+        lq_latent_unpatchify_factor: int = 1,
+        lq_conv_padding_mode: str = "zeros",
+        lq_gate_per_token: bool = False,
+        pit_lq_inject: bool = False,
         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,
@@ -165,6 +185,8 @@ class PidNet(PixDiT_T2I):
         for blk in self.pixel_blocks:
             blk._rope_fn = _pit_rope_fn
 
+        self.pit_lq_inject = pit_lq_inject
+
         num_lq_outputs = (self.patch_depth + lq_interval - 1) // lq_interval
         self.lq_proj = LQProjection2D(
             latent_channels=lq_latent_channels,
@@ -173,13 +195,20 @@ class PidNet(PixDiT_T2I):
             patch_size=self.patch_size,
             sr_scale=sr_scale,
             latent_spatial_down_factor=latent_spatial_down_factor,
+            latent_unpatchify_factor=lq_latent_unpatchify_factor,
             num_res_blocks=lq_num_res_blocks,
             num_outputs=num_lq_outputs,
             interval=lq_interval,
+            conv_padding_mode=lq_conv_padding_mode,
+            gate_per_token=lq_gate_per_token,
+            pit_output=pit_lq_inject,
             dtype=dtype,
             device=device,
             operations=operations,
         )
+        self.pit_lq_gate = SigmaAwareGate(
+            self.hidden_size, per_token=lq_gate_per_token, dtype=dtype, device=device, operations=operations
+        ) if pit_lq_inject else None
 
     def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts):
         return precompute_freqs_cis_2d(
@@ -197,6 +226,11 @@ class PidNet(PixDiT_T2I):
             return s
         return self.lq_proj.gate(s, pid_lq_features[out_idx], pid_degrade_sigma, out_idx)
 
+    def _pre_pixel_blocks(self, s, pid_pit_lq_feature=None, pid_degrade_sigma=None, **kwargs):
+        if pid_pit_lq_feature is None:
+            return s
+        return self.pit_lq_gate(s, pid_pit_lq_feature, pid_degrade_sigma)
+
     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")
@@ -216,12 +250,14 @@ class PidNet(PixDiT_T2I):
             degrade_sigma = degrade_sigma.expand(B).contiguous()
 
         lq_features = self.lq_proj(lq_latent=lq_latent.to(x), target_pH=Hs, target_pW=Ws)
+        pit_lq_feature = lq_features.pop() if self.pit_lq_inject else None
 
         return super()._forward(
             x, timesteps,
             context=context, attention_mask=attention_mask,
             transformer_options=transformer_options,
             pid_lq_features=lq_features,
+            pid_pit_lq_feature=pit_lq_feature,
             pid_degrade_sigma=degrade_sigma,
             **kwargs,
         )

comfy/ldm/seedvr/attention.py

@@ -0,0 +1,51 @@
+import torch
+
+from comfy.ldm.modules import attention as _attention
+
+
+def _var_attention_qkv(q, k, v, heads, skip_reshape):
+    if skip_reshape:
+        return q, k, v, q.shape[-1]
+    total_tokens, embed_dim = q.shape
+    head_dim = embed_dim // heads
+    return (
+        q.view(total_tokens, heads, head_dim),
+        k.view(k.shape[0], heads, head_dim),
+        v.view(v.shape[0], heads, head_dim),
+        head_dim,
+    )
+
+
+def _var_attention_output(out, heads, head_dim, skip_output_reshape):
+    if skip_output_reshape:
+        return out
+    return out.reshape(-1, heads * head_dim)
+
+
+def var_attention_optimized_split(q, k, v, heads, cu_seqlens_q, cu_seqlens_k, *args, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    q, k, v, head_dim = _var_attention_qkv(q, k, v, heads, skip_reshape)
+
+    q_split_indices = cu_seqlens_q[1:-1]
+    k_split_indices = cu_seqlens_k[1:-1]
+    if k.shape[0] != v.shape[0]:
+        raise ValueError("cu_seqlens_k does not match v token count")
+
+    q_splits = torch.tensor_split(q, q_split_indices, dim=0)
+    k_splits = torch.tensor_split(k, k_split_indices, dim=0)
+    v_splits = torch.tensor_split(v, k_split_indices, dim=0)
+    if len(q_splits) != len(k_splits) or len(q_splits) != len(v_splits):
+        raise ValueError("cu_seqlens_q and cu_seqlens_k must describe the same sequence count")
+
+    out = []
+    for q_i, k_i, v_i in zip(q_splits, k_splits, v_splits):
+        q_i = q_i.permute(1, 0, 2).unsqueeze(0)
+        k_i = k_i.permute(1, 0, 2).unsqueeze(0)
+        v_i = v_i.permute(1, 0, 2).unsqueeze(0)
+        out_i = _attention.optimized_attention(q_i, k_i, v_i, heads, skip_reshape=True, skip_output_reshape=True)
+        out.append(out_i.squeeze(0).permute(1, 0, 2))
+
+    out = torch.cat(out, dim=0)
+    return _var_attention_output(out, heads, head_dim, skip_output_reshape)
+
+
+optimized_var_attention = var_attention_optimized_split

comfy/ldm/seedvr/color_fix.py

@@ -0,0 +1,301 @@
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+
+from comfy.ldm.seedvr.constants import (
+    CIELAB_DELTA,
+    CIELAB_KAPPA,
+    D65_WHITE_X,
+    D65_WHITE_Z,
+    WAVELET_DECOMP_LEVELS,
+)
+
+
+def wavelet_blur(image: Tensor, radius):
+    max_safe_radius = max(1, min(image.shape[-2:]) // 8)
+    if radius > max_safe_radius:
+        radius = max_safe_radius
+
+    num_channels = image.shape[1]
+
+    kernel_vals = [
+        [0.0625, 0.125, 0.0625],
+        [0.125,  0.25,  0.125],
+        [0.0625, 0.125, 0.0625],
+    ]
+    kernel = torch.tensor(kernel_vals, dtype=image.dtype, device=image.device)
+    kernel = kernel[None, None].repeat(num_channels, 1, 1, 1)
+
+    image = F.pad(image, (radius, radius, radius, radius), mode='replicate')
+    output = F.conv2d(image, kernel, groups=num_channels, dilation=radius)
+
+    return output
+
+def wavelet_decomposition(image: Tensor, levels: int = WAVELET_DECOMP_LEVELS):
+    high_freq = torch.zeros_like(image)
+
+    for i in range(levels):
+        radius = 2 ** i
+        low_freq = wavelet_blur(image, radius)
+        high_freq.add_(image).sub_(low_freq)
+        image = low_freq
+
+    return high_freq, low_freq
+
+def wavelet_reconstruction(content_feat: Tensor, style_feat: Tensor) -> Tensor:
+
+    if content_feat.shape != style_feat.shape:
+        if len(content_feat.shape) >= 3:
+            style_feat = F.interpolate(
+                style_feat,
+                size=content_feat.shape[-2:],
+                mode='bilinear',
+                align_corners=False
+            )
+
+    content_high_freq, content_low_freq = wavelet_decomposition(content_feat)
+    del content_low_freq
+
+    style_high_freq, style_low_freq = wavelet_decomposition(style_feat)
+    del style_high_freq
+
+    if content_high_freq.shape != style_low_freq.shape:
+        style_low_freq = F.interpolate(
+            style_low_freq,
+            size=content_high_freq.shape[-2:],
+            mode='bilinear',
+            align_corners=False
+        )
+
+    content_high_freq.add_(style_low_freq)
+
+    return content_high_freq.clamp_(-1.0, 1.0)
+
+def _histogram_matching_channel(source: Tensor, reference: Tensor) -> Tensor:
+    original_shape = source.shape
+
+    source_flat = source.flatten()
+    reference_flat = reference.flatten()
+
+    source_sorted, source_indices = torch.sort(source_flat)
+    reference_sorted, _ = torch.sort(reference_flat)
+    del reference_flat
+
+    n_source = len(source_sorted)
+    n_reference = len(reference_sorted)
+
+    if n_source == n_reference:
+        matched_sorted = reference_sorted
+    else:
+        source_quantiles = torch.linspace(0, 1, n_source, device=source.device)
+        ref_indices = (source_quantiles * (n_reference - 1)).long()
+        ref_indices.clamp_(0, n_reference - 1)
+        matched_sorted = reference_sorted[ref_indices]
+        del source_quantiles, ref_indices, reference_sorted
+
+    del source_sorted, source_flat
+
+    inverse_indices = torch.argsort(source_indices)
+    del source_indices
+    matched_flat = matched_sorted[inverse_indices]
+    del matched_sorted, inverse_indices
+
+    return matched_flat.reshape(original_shape)
+
+def _lab_to_rgb_batch(lab: Tensor, matrix_inv: Tensor, epsilon: float, kappa: float) -> Tensor:
+    L, a, b = lab[:, 0], lab[:, 1], lab[:, 2]
+
+    fy = (L + 16.0) / 116.0
+    fx = a.div(500.0).add_(fy)
+    fz = fy - b / 200.0
+    del L, a, b
+
+    x = torch.where(
+        fx > epsilon,
+        torch.pow(fx, 3.0),
+        fx.mul(116.0).sub_(16.0).div_(kappa)
+    )
+    y = torch.where(
+        fy > epsilon,
+        torch.pow(fy, 3.0),
+        fy.mul(116.0).sub_(16.0).div_(kappa)
+    )
+    z = torch.where(
+        fz > epsilon,
+        torch.pow(fz, 3.0),
+        fz.mul(116.0).sub_(16.0).div_(kappa)
+    )
+    del fx, fy, fz
+
+    x.mul_(D65_WHITE_X)
+    z.mul_(D65_WHITE_Z)
+
+    xyz = torch.stack([x, y, z], dim=1)
+    del x, y, z
+
+    B, _, H, W = xyz.shape
+    xyz_flat = xyz.permute(0, 2, 3, 1).reshape(-1, 3)
+    del xyz
+
+    xyz_flat = xyz_flat.to(dtype=matrix_inv.dtype)
+    rgb_linear_flat = torch.matmul(xyz_flat, matrix_inv.T)
+    del xyz_flat
+
+    rgb_linear = rgb_linear_flat.reshape(B, H, W, 3).permute(0, 3, 1, 2)
+    del rgb_linear_flat
+
+    mask = rgb_linear > 0.0031308
+    rgb = torch.where(
+        mask,
+        torch.pow(torch.clamp(rgb_linear, min=0.0), 1.0 / 2.4).mul_(1.055).sub_(0.055),
+        rgb_linear * 12.92
+    )
+    del mask, rgb_linear
+
+    return torch.clamp(rgb, 0.0, 1.0)
+
+def _rgb_to_lab_batch(rgb: Tensor, matrix: Tensor, epsilon: float, kappa: float) -> Tensor:
+    mask = rgb > 0.04045
+    rgb_linear = torch.where(
+        mask,
+        torch.pow((rgb + 0.055) / 1.055, 2.4),
+        rgb / 12.92
+    )
+    del mask
+
+    B, _, H, W = rgb_linear.shape
+    rgb_flat = rgb_linear.permute(0, 2, 3, 1).reshape(-1, 3)
+    del rgb_linear
+
+    rgb_flat = rgb_flat.to(dtype=matrix.dtype)
+    xyz_flat = torch.matmul(rgb_flat, matrix.T)
+    del rgb_flat
+
+    xyz = xyz_flat.reshape(B, H, W, 3).permute(0, 3, 1, 2)
+    del xyz_flat
+
+    xyz[:, 0].div_(D65_WHITE_X)
+    xyz[:, 2].div_(D65_WHITE_Z)
+
+    epsilon_cubed = epsilon ** 3
+    mask = xyz > epsilon_cubed
+    f_xyz = torch.where(
+        mask,
+        torch.pow(xyz, 1.0 / 3.0),
+        xyz.mul(kappa).add_(16.0).div_(116.0)
+    )
+    del xyz, mask
+
+    L = f_xyz[:, 1].mul(116.0).sub_(16.0)
+    a = (f_xyz[:, 0] - f_xyz[:, 1]).mul_(500.0)
+    b = (f_xyz[:, 1] - f_xyz[:, 2]).mul_(200.0)
+    del f_xyz
+
+    return torch.stack([L, a, b], dim=1)
+
+def lab_color_transfer(
+    content_feat: Tensor,
+    style_feat: Tensor,
+    luminance_weight: float = 0.8
+) -> Tensor:
+    content_feat = wavelet_reconstruction(content_feat, style_feat)
+
+    if content_feat.shape != style_feat.shape:
+        style_feat = F.interpolate(
+            style_feat,
+            size=content_feat.shape[-2:],
+            mode='bilinear',
+            align_corners=False
+        )
+
+    device = content_feat.device
+    original_dtype = content_feat.dtype
+    content_feat = content_feat.float()
+    style_feat = style_feat.float()
+
+    rgb_to_xyz_matrix = torch.tensor([
+        [0.4124564, 0.3575761, 0.1804375],
+        [0.2126729, 0.7151522, 0.0721750],
+        [0.0193339, 0.1191920, 0.9503041]
+    ], dtype=torch.float32, device=device)
+
+    xyz_to_rgb_matrix = torch.tensor([
+        [ 3.2404542, -1.5371385, -0.4985314],
+        [-0.9692660,  1.8760108,  0.0415560],
+        [ 0.0556434, -0.2040259,  1.0572252]
+    ], dtype=torch.float32, device=device)
+
+    epsilon = CIELAB_DELTA
+    kappa = CIELAB_KAPPA
+
+    content_feat.add_(1.0).mul_(0.5).clamp_(0.0, 1.0)
+    style_feat.add_(1.0).mul_(0.5).clamp_(0.0, 1.0)
+
+    content_lab = _rgb_to_lab_batch(content_feat, rgb_to_xyz_matrix, epsilon, kappa)
+    del content_feat
+
+    style_lab = _rgb_to_lab_batch(style_feat, rgb_to_xyz_matrix, epsilon, kappa)
+    del style_feat, rgb_to_xyz_matrix
+
+    matched_a = _histogram_matching_channel(content_lab[:, 1], style_lab[:, 1])
+    matched_b = _histogram_matching_channel(content_lab[:, 2], style_lab[:, 2])
+
+    if luminance_weight < 1.0:
+        matched_L = _histogram_matching_channel(content_lab[:, 0], style_lab[:, 0])
+        result_L = content_lab[:, 0].mul(luminance_weight).add_(matched_L.mul(1.0 - luminance_weight))
+        del matched_L
+    else:
+        result_L = content_lab[:, 0]
+
+    del content_lab, style_lab
+
+    result_lab = torch.stack([result_L, matched_a, matched_b], dim=1)
+    del result_L, matched_a, matched_b
+
+    result_rgb = _lab_to_rgb_batch(result_lab, xyz_to_rgb_matrix, epsilon, kappa)
+    del result_lab, xyz_to_rgb_matrix
+
+    result = result_rgb.mul_(2.0).sub_(1.0)
+    del result_rgb
+
+    result = result.to(original_dtype)
+
+    return result
+
+
+def wavelet_color_transfer(content_feat: Tensor, style_feat: Tensor) -> Tensor:
+    return wavelet_reconstruction(content_feat, style_feat)
+
+
+def adain_color_transfer(content_feat: Tensor, style_feat: Tensor, eps: float = 1e-5) -> Tensor:
+    if content_feat.shape != style_feat.shape:
+        style_feat = F.interpolate(
+            style_feat,
+            size=content_feat.shape[-2:],
+            mode='bilinear',
+            align_corners=False,
+        )
+
+    original_dtype = content_feat.dtype
+    content_feat = content_feat.float()
+    style_feat = style_feat.float()
+
+    b, c = content_feat.shape[:2]
+    content_flat = content_feat.reshape(b, c, -1)
+    style_flat = style_feat.reshape(b, c, -1)
+
+    content_mean = content_flat.mean(dim=2).reshape(b, c, 1, 1)
+    content_std = (content_flat.var(dim=2, correction=0) + eps).sqrt().reshape(b, c, 1, 1)
+    style_mean = style_flat.mean(dim=2).reshape(b, c, 1, 1)
+    style_std = (style_flat.var(dim=2, correction=0) + eps).sqrt().reshape(b, c, 1, 1)
+    del content_flat, style_flat
+
+    normalized = (content_feat - content_mean) / content_std
+    del content_mean, content_std
+    result = normalized * style_std + style_mean
+    del normalized, style_mean, style_std
+
+    result = result.clamp_(-1.0, 1.0)
+    if result.dtype != original_dtype:
+        result = result.to(original_dtype)
+    return result

comfy/ldm/seedvr/constants.py

@@ -0,0 +1,48 @@
+"""SeedVR2 constants."""
+
+# Temporal chunk-size law: the sampler's activation wall is linear in
+# T_latent * pixel area (17-cell resolution sweep + T bisection, RTX 5090, 3b fp16):
+#   max_latent_frames = (free_GiB - RESERVED - K*SIGMA) / (GIB_PER_MPX_FRAME * megapixels)
+# RESERVED covers model staging plus fixed CUDA/torch overhead; SIGMA is the measured
+# run-to-run spread of the wall; K=4 trades ~10% smaller chunks for ~1e-5 OOM odds.
+SEEDVR2_CHUNK_GIB_PER_MPX_FRAME = 0.55
+SEEDVR2_CHUNK_RESERVED_GIB = 8.5
+SEEDVR2_CHUNK_SIGMA_GIB = 0.55
+SEEDVR2_CHUNK_SIGMA_K = 4
+
+SEEDVR2_7B_VID_DIM = 3072
+SEEDVR2_OOM_BACKOFF_DIVISOR = 2
+SEEDVR2_DTYPE_BYTES_FLOOR = 4
+SEEDVR2_7B_MLP_CHUNK = 8192
+SEEDVR2_ROPE_PARTIAL_CHUNK_TOKENS = 4096  # partial-RoPE application token-chunk.
+SEEDVR2_LATENT_CHANNELS = 16
+
+SEEDVR2_COLOR_MEM_HEADROOM = 0.75
+SEEDVR2_LAB_SCALE_MULTIPLIER = 13
+SEEDVR2_WAVELET_SCALE_MULTIPLIER = 10  # per-frame byte multiplier, wavelet path.
+SEEDVR2_ADAIN_SCALE_MULTIPLIER = 6
+
+BYTEDANCE_VAE_SCALING_FACTOR = 0.9152   # configs_3b/main.yaml:57.
+BYTEDANCE_VAE_SHIFTING_FACTOR = 0.0
+BYTEDANCE_VAE_CONV_MEM_GIB = 0.5
+BYTEDANCE_VAE_NORM_MEM_GIB = 0.5
+BYTEDANCE_LOGVAR_CLAMP_MIN = -30.0      # video_vae_v3/modules/types.py:28.
+BYTEDANCE_LOGVAR_CLAMP_MAX = 20.0       # video_vae_v3/modules/types.py:28.
+BYTEDANCE_GN_CHUNKS_FP16 = 4            # causal_inflation_lib.py:351 (GroupNorm chunk count, fp16).
+BYTEDANCE_GN_CHUNKS_FP32 = 2            # causal_inflation_lib.py:351 (GroupNorm chunk count, fp32).
+BYTEDANCE_BLOCK_OUT_CHANNELS = (128, 256, 512, 512)  # s8_c16_t4_inflation_sd3.yaml:7-11.
+BYTEDANCE_SLICING_SAMPLE_MIN = 4        # s8_c16_t4_inflation_sd3.yaml:22 (slicing_sample_min_size).
+BYTEDANCE_VAE_TEMPORAL_DOWNSAMPLE = 4   # infer.py:230 (temporal_downsample_factor); the 4n+1 factor.
+BYTEDANCE_VAE_SPATIAL_DOWNSAMPLE = 8    # infer.py:231 (spatial_downsample_factor).
+BYTEDANCE_720P_REF_AREA = 45 * 80       # dit_v2/window.py:32 (720p reference area for window scaling).
+BYTEDANCE_MAX_TEMPORAL_WINDOW = 30      # dit_v2/window.py:35 (max temporal window frames).
+BYTEDANCE_ROPE_MAX_FREQ = 256           # dit_v2/rope.py:31 (pixel-RoPE max frequency).
+BYTEDANCE_SINUSOIDAL_DIM = 256          # dit_3b/nadit.py:120 (timestep sinusoidal embed dim).
+
+ROPE_THETA = 10000   # RoPE base; Su et al., "RoFormer", arXiv:2104.09864.
+
+CIELAB_DELTA = 6.0 / 29.0          # CIE 15 (delta).
+CIELAB_KAPPA = (29.0 / 3.0) ** 3   # CIE 15 (kappa).
+D65_WHITE_X = 0.95047              # CIE D65 standard illuminant Xn (Yn = 1).
+D65_WHITE_Z = 1.08883              # CIE D65 standard illuminant Zn.
+WAVELET_DECOMP_LEVELS = 5          # wavelet color-fix decomposition depth (GIMP/Krita; StableSR).

comfy/model_base.py

@@ -55,6 +55,7 @@ import comfy.ldm.pixeldit.model
 import comfy.ldm.pixeldit.pid
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
+import comfy.ldm.seedvr.model
 import comfy.ldm.boogu.model
 import comfy.ldm.qwen_image.model
 import comfy.ldm.ideogram4.model
@@ -932,6 +933,17 @@ class HunyuanDiT(BaseModel):
         out['image_meta_size'] = comfy.conds.CONDRegular(torch.FloatTensor([[height, width, target_height, target_width, 0, 0]]))
         return out
 
+class SeedVR2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.seedvr.model.NaDiT)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        condition = kwargs.get("condition", None)
+        if condition is not None:
+            out["condition"] = comfy.conds.CONDRegular(condition)
+        return out
+
 class PixArt(BaseModel):
     def __init__(self, model_config, model_type=ModelType.EPS, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.pixart.pixartms.PixArtMS)

comfy/model_detection.py

@@ -470,15 +470,46 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     # 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])
+        latent_proj_in_channels = int(state_dict[_lq_w_key].shape[1])
+        hidden_dim = int(state_dict[_lq_w_key].shape[0])
         _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)})
+        pid_v1_5 = '{}lq_proj.pit_head.weight'.format(key_prefix) in state_dict_keys
         dit_config = {"image_model": "pid",
-                      "lq_latent_channels": in_ch,
+                      "lq_hidden_dim": hidden_dim}
-                      "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
+        if pid_v1_5:
+            pid_v1_5_variants = {
+                16: {  # Flux and QwenImage
+                    "lq_latent_channels": 16,
+                    "latent_spatial_down_factor": 8,
+                    "lq_latent_unpatchify_factor": 1,
+                },
+                32: {  # Flux2 after 2x latent unpatchify
+                    "lq_latent_channels": 128,
+                    "latent_spatial_down_factor": 16,
+                    "lq_latent_unpatchify_factor": 2,
+                },
+            }
+            variant = pid_v1_5_variants.get(latent_proj_in_channels)
+            if variant is None:
+                raise ValueError(f"Unsupported PiD v1.5 latent projection with {latent_proj_in_channels} input channels")
+            gate_weight = state_dict['{}lq_proj.gate_modules.0.content_proj.weight'.format(key_prefix)]
+            dit_config.update(variant)
+            dit_config.update({
+                "lq_conv_padding_mode": "replicate",
+                "lq_gate_per_token": gate_weight.shape[0] == 1,
+                "pit_lq_inject": True,
+                "rope_ref_h": 2048,
+                "rope_ref_w": 2048,
+            })
+        else:
+            dit_config.update({
+                "lq_latent_channels": latent_proj_in_channels,
+                "latent_spatial_down_factor": 16 if latent_proj_in_channels >= 64 else 8,
+            })
         return dit_config
 
     if '{}core.pixel_embedder.proj.weight'.format(key_prefix) in state_dict_keys:  # PixelDiT T2I
@@ -598,6 +629,44 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         return dit_config
 
+    seedvr2_7b_separate_key = "{}blocks.35.mlp.vid.proj_out.weight".format(key_prefix)
+    if seedvr2_7b_separate_key in state_dict_keys and state_dict[seedvr2_7b_separate_key].shape[0] == 3072: # seedvr2 7b
+        dit_config = {}
+        dit_config["image_model"] = "seedvr2"
+        dit_config["vid_dim"] = 3072
+        dit_config["heads"] = 24
+        dit_config["num_layers"] = 36
+        # This checkpoint uses separate vid/txt MMModule keys in every block.
+        dit_config["mm_layers"] = 36
+        dit_config["norm_eps"] = 1e-5
+        dit_config["rope_type"] = "rope3d"
+        dit_config["rope_dim"] = 64
+        dit_config["mlp_type"] = "normal"
+        return dit_config
+    if "{}blocks.35.mlp.all.proj_in_gate.weight".format(key_prefix) in state_dict_keys: # seedvr2 7b
+        dit_config = {}
+        dit_config["image_model"] = "seedvr2"
+        dit_config["vid_dim"] = 3072
+        dit_config["heads"] = 24
+        dit_config["num_layers"] = 36
+        # This checkpoint uses shared all.* MMModule keys after the initial blocks.
+        dit_config["mm_layers"] = 10
+        dit_config["norm_eps"] = 1e-5
+        dit_config["rope_type"] = "rope3d"
+        dit_config["rope_dim"] = 64
+        dit_config["mlp_type"] = "swiglu"
+        return dit_config
+    if "{}blocks.31.mlp.all.proj_in_gate.weight".format(key_prefix) in state_dict_keys: # seedvr2 3b
+        dit_config = {}
+        dit_config["image_model"] = "seedvr2"
+        dit_config["vid_dim"] = 2560
+        dit_config["heads"] = 20
+        dit_config["num_layers"] = 32
+        dit_config["norm_eps"] = 1.0e-05
+        dit_config["mlp_type"] = "swiglu"
+        dit_config["vid_out_norm"] = True
+        return dit_config
+
     if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
         dit_config = {}
         dit_config["image_model"] = "wan2.1"
@@ -1119,9 +1188,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
     return unet_config
 
-def model_config_from_unet_config(unet_config, state_dict=None):
+
+def model_config_from_unet_config(unet_config, state_dict=None, unet_key_prefix=""):
     for model_config in comfy.supported_models.models:
-        if model_config.matches(unet_config, state_dict):
+        if model_config.matches(unet_config, state_dict, unet_key_prefix=unet_key_prefix):
             return model_config(unet_config)
 
     logging.error("no match {}".format(unet_config))
@@ -1131,7 +1201,7 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
     unet_config = detect_unet_config(state_dict, unet_key_prefix, metadata=metadata)
     if unet_config is None:
         return None
-    model_config = model_config_from_unet_config(unet_config, state_dict)
+    model_config = model_config_from_unet_config(unet_config, state_dict, unet_key_prefix)
     if model_config is None and use_base_if_no_match:
         model_config = comfy.supported_models_base.BASE(unet_config)
 

comfy/model_management.py

@@ -616,6 +616,8 @@ PIN_PRESSURE_HYSTERESIS = 256 * 1024 * 1024
 #Freeing registerables on pressure does imply a GPU sync, so go big on
 #the hysteresis so each expensive sync gives us back a good chunk.
 REGISTERABLE_PIN_HYSTERESIS = 2048 * 1024 * 1024
+WINDOWS_PIN_EVICTION_SWAP_PERCENT = 5.0
+WINDOWS_PIN_EVICTION_EMERGENCY_AVAILABLE = 512 * 1024 ** 2
 
 def module_size(module):
     module_mem = 0
@@ -642,6 +644,15 @@ def free_pins(size, evict_active=False):
             size -= freed
     return freed_total
 
+def should_free_pins_for_ram_pressure(shortfall):
+    if shortfall <= 0:
+        return False
+    if not WINDOWS:
+        return True
+    if psutil.virtual_memory().available < WINDOWS_PIN_EVICTION_EMERGENCY_AVAILABLE:
+        return True
+    return psutil.swap_memory().percent >= WINDOWS_PIN_EVICTION_SWAP_PERCENT
+
 def ensure_pin_budget(size, evict_active=False):
     if args.high_ram:
         return True

comfy/ops.py

@@ -1104,6 +1104,21 @@ def _load_quantized_module(module, super_load, state_dict, prefix, local_metadat
                 scales["convrot_groupsize"] = int(
                     layer_conf.get("convrot_groupsize", params_conf.get("convrot_groupsize", 256))
                 )
+        elif module.quant_format == "convrot_w4a4":
+            scale = pop_scale("weight_scale")
+            if scale is None:
+                raise ValueError(f"Missing ConvRot W4A4 weight scale for layer {layer_name}")
+            params_conf = layer_conf.get("params", {})
+            if not isinstance(params_conf, dict):
+                params_conf = {}
+            scales = {
+                "scale": scale,
+                "convrot_groupsize": int(
+                    layer_conf.get("convrot_groupsize", params_conf.get("convrot_groupsize", 256))
+                ),
+                "quant_group_size": 64,
+                "linear_dtype": layer_conf.get("linear_dtype", params_conf.get("linear_dtype", "int4")),
+            }
         else:
             raise ValueError(f"Unsupported quantization format: {module.quant_format}")
 
@@ -1150,6 +1165,11 @@ def _quantized_weight_state_dict(module, sd, prefix, extra_quant_conf=None, extr
         if module.quant_format == "int8_tensorwise" and getattr(params, "convrot", False):
             quant_conf["convrot"] = True
             quant_conf["convrot_groupsize"] = getattr(params, "convrot_groupsize", 256)
+        elif module.quant_format == "convrot_w4a4":
+            quant_conf["convrot_groupsize"] = getattr(params, "convrot_groupsize", 256)
+            linear_dtype = getattr(params, "linear_dtype", "int4")
+            if linear_dtype != "int4":
+                quant_conf["linear_dtype"] = linear_dtype
         if extra_quant_conf:
             quant_conf.update(extra_quant_conf)
         sd[f"{prefix}comfy_quant"] = torch.tensor(list(json.dumps(quant_conf).encode("utf-8")), dtype=torch.uint8)
@@ -1237,7 +1257,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 run_every_op()
 
                 input_shape = input.shape
-                reshaped_3d = False
+                reshaped_nd = False
                 #If cast needs to apply lora, it should be done in the compute dtype
                 compute_dtype = input.dtype
 
@@ -1274,12 +1294,12 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 # Inference path (unchanged)
                 if _use_quantized and quantize_input:
 
-                    # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
+                    # Reshape >=3D tensors to 2D for quantization (needed for NVFP4 and others)
-                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
+                    input_reshaped = input.reshape(-1, input_shape[-1]) if input.ndim >= 3 else input
 
                     # Fall back to non-quantized for non-2D tensors
                     if input_reshaped.ndim == 2:
-                        reshaped_3d = input.ndim == 3
+                        reshaped_nd = input.ndim >= 3
                         # dtype is now implicit in the layout class
                         scale = getattr(self, 'input_scale', None)
                         if scale is not None:
@@ -1294,9 +1314,9 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                     weight_only_quant=weight_only_quant,
                 )
 
-                # Reshape output back to 3D if input was 3D
+                # Reshape output back to original rank if input was >2D
-                if reshaped_3d:
+                if reshaped_nd:
-                    output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
+                    output = output.reshape((*input_shape[:-1], self.weight.shape[0]))
 
                 return output
 
@@ -1430,6 +1450,12 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 }
                 if hasattr(params, "block_scale"): # NVFP4
                     kwargs["block_scale"] = params.block_scale[i]
+                if hasattr(params, "quant_group_size"):
+                    kwargs["quant_group_size"] = params.quant_group_size
+                if hasattr(params, "convrot_groupsize"):
+                    kwargs["convrot_groupsize"] = params.convrot_groupsize
+                if hasattr(params, "linear_dtype"):
+                    kwargs["linear_dtype"] = params.linear_dtype
                 return QuantizedTensor(weight._qdata[i], weight._layout_cls, type(params)(**kwargs))
 
             def state_dict(self, *args, destination=None, prefix="", **kwargs):

comfy/quant_ops.py

@@ -3,6 +3,22 @@ import logging
 
 from comfy.cli_args import args
 
+
+def _rocm_kitchen_arch_supported():
+    """comfy-kitchen's INT8 Triton kernels compile tl.dot to matrix-core instructions.
+    RDNA3/3.5/4 (gfx11xx/gfx12xx) have WMMA and CDNA (gfx9xx) has MFMA; RDNA1/RDNA2
+    (gfx10xx) have neither, so the INT8 path hangs the GPU there. Gates the automatic
+    ROCm default so those cards stay on the eager fallback (an explicit
+    --enable-triton-backend still forces it on any arch)."""
+    try:
+        arch = torch.cuda.get_device_properties(torch.cuda.current_device()).gcnArchName.split(":")[0]
+    except Exception:
+        return False
+    if arch.startswith(("gfx11", "gfx12")):
+        return True
+    return arch in ("gfx908", "gfx90a", "gfx940", "gfx941", "gfx942", "gfx950")
+
+
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -10,6 +26,7 @@ try:
         QuantizedLayout,
         TensorCoreFP8Layout as _CKFp8Layout,
         TensorCoreNVFP4Layout as _CKNvfp4Layout,
+        TensorCoreConvRotW4A4Layout as _CKTensorCoreConvRotW4A4Layout,
         TensorWiseINT8Layout as _CKTensorWiseINT8Layout,
         register_layout_op,
         register_layout_class,
@@ -24,10 +41,22 @@ try:
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
-    if args.enable_triton_backend:
+    # On ROCm/AMD the CUDA backend is unavailable, so Triton is the only accelerated
+    # comfy-kitchen backend. Enable it by default there, but only on Triton >= 3.7 AND a
+    # matrix-core GPU (RDNA3+ WMMA gfx11xx/gfx12xx, CDNA MFMA gfx9xx). RDNA1/RDNA2
+    # (gfx10xx) have no WMMA -> the INT8 tl.dot path hangs the GPU, so they stay eager.
+    # older Triton lacks libdevice.rint on the HIP backend and hard-crashes the INT8 path.
+    if args.disable_triton_backend:
+        ck.registry.disable("triton")
+    elif args.enable_triton_backend: # or (torch.version.hip is not None and _rocm_kitchen_arch_supported()):
         try:
             import triton
-            logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+            triton_version = tuple(int(v) for v in triton.__version__.split(".")[:2])
+            if args.enable_triton_backend or triton_version >= (3, 7):
+                logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+            else:
+                logging.info("Triton %s is too old for the ROCm INT8 path (needs >= 3.7); comfy-kitchen triton backend disabled.", triton.__version__)
+                ck.registry.disable("triton")
         except ImportError as e:
             logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
             ck.registry.disable("triton")
@@ -51,6 +80,9 @@ except ImportError as e:
     class _CKTensorWiseINT8Layout:
         pass
 
+    class _CKTensorCoreConvRotW4A4Layout:
+        pass
+
     def register_layout_class(name, cls):
         pass
 
@@ -179,6 +211,7 @@ class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase):
 # Backward compatibility alias - default to E4M3
 TensorCoreFP8Layout = TensorCoreFP8E4M3Layout
 TensorWiseINT8Layout = _CKTensorWiseINT8Layout
+TensorCoreConvRotW4A4Layout = _CKTensorCoreConvRotW4A4Layout
 
 
 # ==============================================================================
@@ -190,6 +223,7 @@ register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
 register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
 register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
 register_layout_class("TensorWiseINT8Layout", _CKTensorWiseINT8Layout)
+register_layout_class("TensorCoreConvRotW4A4Layout", _CKTensorCoreConvRotW4A4Layout)
 if _CK_MXFP8_AVAILABLE:
     register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout)
 
@@ -227,6 +261,13 @@ QUANT_ALGOS["int8_tensorwise"] = {
     "quantize_input": False,
 }
 
+QUANT_ALGOS["convrot_w4a4"] = {
+    "storage_t": torch.int8,
+    "parameters": {"weight_scale"},
+    "comfy_tensor_layout": "TensorCoreConvRotW4A4Layout",
+    "quantize_input": False,
+}
+
 
 # ==============================================================================
 # Re-exports for backward compatibility
@@ -239,6 +280,7 @@ __all__ = [
     "TensorCoreFP8E4M3Layout",
     "TensorCoreFP8E5M2Layout",
     "TensorCoreNVFP4Layout",
+    "TensorCoreConvRotW4A4Layout",
     "TensorWiseINT8Layout",
     "QUANT_ALGOS",
     "register_layout_op",

comfy/sd.py

@@ -16,6 +16,7 @@ import comfy.ldm.cosmos.vae
 import comfy.ldm.wan.vae
 import comfy.ldm.wan.vae2_2
 import comfy.ldm.hunyuan3d.vae
+import comfy.ldm.seedvr.vae
 import comfy.ldm.triposplat.vae
 import comfy.ldm.ace.vae.music_dcae_pipeline
 import comfy.ldm.cogvideo.vae
@@ -473,7 +474,8 @@ class CLIP:
 
 class VAE:
     def __init__(self, sd=None, device=None, config=None, dtype=None, metadata=None):
-        if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
+        is_seedvr2_vae = "decoder.up_blocks.2.upsamplers.0.upscale_conv.weight" in sd
+        if not is_seedvr2_vae and 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
             sd = diffusers_convert.convert_vae_state_dict(sd)
 
         if model_management.is_amd():
@@ -500,6 +502,8 @@ class VAE:
         self.upscale_index_formula = None
         self.extra_1d_channel = None
         self.crop_input = True
+        self.handles_tiling = False
+        self.format_encoded = None
 
         self.audio_sample_rate = 44100
 
@@ -546,6 +550,22 @@ class VAE:
                 self.first_stage_model = StageC_coder()
                 self.downscale_ratio = 32
                 self.latent_channels = 16
+            elif "decoder.up_blocks.2.upsamplers.0.upscale_conv.weight" in sd: # seedvr2
+                self.first_stage_model = comfy.ldm.seedvr.vae.VideoAutoencoderKLWrapper()
+                self.latent_channels = comfy.ldm.seedvr.vae.SEEDVR2_LATENT_CHANNELS
+                self.latent_dim = 3
+                self.disable_offload = True
+                self.memory_used_decode = lambda shape, dtype: self.first_stage_model.comfy_memory_used_decode(shape)
+                self.memory_used_encode = lambda shape, dtype: (max(shape[2], 5) * shape[3] * shape[4] * 64) * model_management.dtype_size(dtype)
+                self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+                self.handles_tiling = True
+                self.format_encoded = self.first_stage_model.comfy_format_encoded
+                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                self.downscale_index_formula = (4, 8, 8)
+                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                self.upscale_index_formula = (4, 8, 8)
+                self.process_input = lambda image: image * 2.0 - 1.0
+                self.crop_input = False
             elif "decoder.conv_in.weight" in sd:
                 if sd['decoder.conv_in.weight'].shape[1] == 64:
                     ddconfig = {"block_out_channels": [128, 256, 512, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 32, "downsample_match_channel": True, "upsample_match_channel": True}
@@ -1012,6 +1032,10 @@ class VAE:
         decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, index_formulas=self.upscale_index_formula, output_device=self.output_device))
 
+    def _decode_tiled_owned(self, samples, **kwargs):
+        out = self.first_stage_model.decode_tiled(samples.to(self.vae_dtype).to(self.device), **kwargs)
+        return self.process_output(out.to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True))
+
     def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
         steps = pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x, tile_y, overlap)
         steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x // 2, tile_y * 2, overlap)
@@ -1048,6 +1072,25 @@ class VAE:
         encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
+    def _encode_tiled_owned(self, pixel_samples, **kwargs):
+        x = self.process_input(pixel_samples).to(self.vae_dtype).to(self.device)
+        out = self.first_stage_model.encode_tiled(x, **kwargs)
+        return out.to(device=self.output_device, dtype=self.vae_output_dtype())
+
+    def _owned_tiled_args(self, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
+        args = {}
+        if tile_x is not None:
+            args["tile_x"] = tile_x
+        if tile_y is not None:
+            args["tile_y"] = tile_y
+        if overlap is not None:
+            args["overlap"] = overlap
+        if tile_t is not None:
+            args["tile_t"] = tile_t
+        if overlap_t is not None:
+            args["overlap_t"] = overlap_t
+        return args
+
     def decode(self, samples_in, vae_options={}):
         self.throw_exception_if_invalid()
         pixel_samples = None
@@ -1095,11 +1138,19 @@ class VAE:
                 if dims == 1 or self.extra_1d_channel is not None:
                     pixel_samples = self.decode_tiled_1d(samples_in)
                 elif dims == 2:
-                    pixel_samples = self.decode_tiled_(samples_in)
+                    if self.handles_tiling:
+                        tile = 256 // self.spacial_compression_decode()
+                        overlap = tile // 4
+                        pixel_samples = self._decode_tiled_owned(samples_in, tile_x=tile, tile_y=tile, overlap=overlap)
+                    else:
+                        pixel_samples = self.decode_tiled_(samples_in)
                 elif dims == 3:
                     tile = 256 // self.spacial_compression_decode()
                     overlap = tile // 4
-                    pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
+                    if self.handles_tiling:
+                        pixel_samples = self._decode_tiled_owned(samples_in, tile_x=tile, tile_y=tile, overlap=overlap)
+                    else:
+                        pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
 
         pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1)
         return pixel_samples
@@ -1118,7 +1169,9 @@ class VAE:
             args["overlap"] = overlap
 
         with model_management.cuda_device_context(self.device):
-            if dims == 1 or self.extra_1d_channel is not None:
+            if self.handles_tiling and dims in (2, 3):
+                output = self._decode_tiled_owned(samples, **self._owned_tiled_args(tile_x, tile_y, overlap, tile_t, overlap_t))
+            elif dims == 1 or self.extra_1d_channel is not None:
                 args.pop("tile_y")
                 output = self.decode_tiled_1d(samples, **args)
             elif dims == 2:
@@ -1179,12 +1232,17 @@ class VAE:
                 if self.latent_dim == 3:
                     tile = 256
                     overlap = tile // 4
-                    samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
+                    if self.handles_tiling:
+                        samples = self._encode_tiled_owned(pixel_samples, tile_x=tile, tile_y=tile, overlap=overlap)
+                    else:
+                        samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
                 elif self.latent_dim == 1 or self.extra_1d_channel is not None:
                     samples = self.encode_tiled_1d(pixel_samples)
                 else:
                     samples = self.encode_tiled_(pixel_samples)
 
+        if self.format_encoded is not None:
+            samples = self.format_encoded(samples)
         return samples
 
     def encode_tiled(self, pixel_samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
@@ -1192,7 +1250,7 @@ class VAE:
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
         dims = self.latent_dim
         pixel_samples = pixel_samples.movedim(-1, 1)
-        if dims == 3:
+        if dims == 3 and pixel_samples.ndim < 5:
             if not self.not_video:
                 pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
             else:
@@ -1216,21 +1274,27 @@ class VAE:
             elif dims == 2:
                 samples = self.encode_tiled_(pixel_samples, **args)
             elif dims == 3:
-                if tile_t is not None:
+                if self.handles_tiling:
-                    tile_t_latent = max(2, self.downscale_ratio[0](tile_t))
+                    samples = self._encode_tiled_owned(pixel_samples, **self._owned_tiled_args(tile_x, tile_y, overlap, tile_t, overlap_t))
                 else:
-                    tile_t_latent = 9999
+                    if tile_t is not None:
-                args["tile_t"] = self.upscale_ratio[0](tile_t_latent)
+                        tile_t_latent = max(2, self.downscale_ratio[0](tile_t))
+                    else:
+                        tile_t_latent = 9999
+                    args["tile_t"] = self.upscale_ratio[0](tile_t_latent)
 
-                if overlap_t is None:
+                    spatial_overlap = overlap if overlap is not None else 64
-                    args["overlap"] = (1, overlap, overlap)
+                    if overlap_t is None:
-                else:
+                        args["overlap"] = (1, spatial_overlap, spatial_overlap)
-                    args["overlap"] = (self.upscale_ratio[0](max(1, min(tile_t_latent // 2, self.downscale_ratio[0](overlap_t)))), overlap, overlap)
+                    else:
-                maximum = pixel_samples.shape[2]
+                        args["overlap"] = (self.upscale_ratio[0](max(1, min(tile_t_latent // 2, self.downscale_ratio[0](overlap_t)))), spatial_overlap, spatial_overlap)
-                maximum = self.upscale_ratio[0](self.downscale_ratio[0](maximum))
+                    maximum = pixel_samples.shape[2]
+                    maximum = self.upscale_ratio[0](self.downscale_ratio[0](maximum))
 
-                samples = self.encode_tiled_3d(pixel_samples[:,:,:maximum], **args)
+                    samples = self.encode_tiled_3d(pixel_samples[:,:,:maximum], **args)
 
+        if self.format_encoded is not None:
+            samples = self.format_encoded(samples)
         return samples
 
     def get_sd(self):
@@ -1898,7 +1962,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
         manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes)
     else:
         manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
-    model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
+    model_config.set_inference_dtype(unet_dtype, manual_cast_dtype, device=load_device)
 
     if model_config.clip_vision_prefix is not None:
         if output_clipvision:
@@ -2039,7 +2103,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable
         manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes)
     else:
         manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
-    model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
+    model_config.set_inference_dtype(unet_dtype, manual_cast_dtype, device=load_device)
 
     if custom_operations is not None:
         model_config.custom_operations = custom_operations

comfy/supported_models.py

@@ -1685,6 +1685,40 @@ class Chroma(supported_models_base.BASE):
         t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.pixart_t5.PixArtTokenizer, comfy.text_encoders.pixart_t5.pixart_te(**t5_detect))
 
+class SeedVR2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "seedvr2"
+    }
+    unet_extra_config = {}
+    required_keys = {
+        "{}positive_conditioning",
+        "{}negative_conditioning",
+    }
+    latent_format = comfy.latent_formats.SeedVR2
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+    sampling_settings = {
+        "shift": 1.0,
+    }
+
+    def set_inference_dtype(self, dtype, manual_cast_dtype, device=None):
+        if (
+            dtype == torch.float16
+            and manual_cast_dtype is None
+            and comfy.model_management.should_use_bf16(device)
+        ):
+            manual_cast_dtype = torch.bfloat16
+        super().set_inference_dtype(dtype, manual_cast_dtype, device=device)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.SeedVR2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None
+
 class ChromaRadiance(Chroma):
     unet_config = {
         "image_model": "chroma_radiance",
@@ -2348,6 +2382,7 @@ models = [
     HiDream,
     HiDreamO1,
     Chroma,
+    SeedVR2,
     ChromaRadiance,
     ACEStep,
     ACEStep15,

comfy/supported_models_base.py

@@ -54,13 +54,13 @@ class BASE:
     optimizations = {"fp8": False}
 
     @classmethod
-    def matches(s, unet_config, state_dict=None):
+    def matches(s, unet_config, state_dict=None, unet_key_prefix=""):
         for k in s.unet_config:
             if k not in unet_config or s.unet_config[k] != unet_config[k]:
                 return False
         if state_dict is not None:
             for k in s.required_keys:
-                if k not in state_dict:
+                if k.format(unet_key_prefix) not in state_dict:
                     return False
         return True
 
@@ -115,7 +115,7 @@ class BASE:
         replace_prefix = {"": self.vae_key_prefix[0]}
         return utils.state_dict_prefix_replace(state_dict, replace_prefix)
 
-    def set_inference_dtype(self, dtype, manual_cast_dtype):
+    def set_inference_dtype(self, dtype, manual_cast_dtype, device=None):
         self.unet_config['dtype'] = dtype
         self.manual_cast_dtype = manual_cast_dtype
 

comfy/text_encoders/gemma4.py

@@ -1088,7 +1088,7 @@ class Gemma4_Tokenizer():
             h, w = samples.shape[2], samples.shape[3]
             patch_size = 16
             pooling_k = 3
-            max_soft_tokens = 70 if is_video else 280  # video uses smaller token budget per frame
+            max_soft_tokens = kwargs.get("max_soft_tokens", 70 if is_video else 280)
             max_patches = max_soft_tokens * pooling_k * pooling_k
             target_px = max_patches * patch_size * patch_size
             factor = (target_px / (h * w)) ** 0.5

comfy_api_nodes/apis/bytedance.py

@@ -17,6 +17,10 @@ class Seedream4Options(BaseModel):
     max_images: int = Field(15)
 
 
+class Seedream5OptimizePromptOptions(BaseModel):
+    thinking: Literal["auto", "enabled", "disabled"] = Field(...)
+
+
 class Seedream4TaskCreationRequest(BaseModel):
     model: str = Field(...)
     prompt: str = Field(...)
@@ -28,6 +32,7 @@ class Seedream4TaskCreationRequest(BaseModel):
     sequential_image_generation_options: Seedream4Options | None = Field(Seedream4Options(max_images=15))
     watermark: bool = Field(False)
     output_format: str | None = None
+    optimize_prompt_options: Seedream5OptimizePromptOptions | None = None
 
 
 class ImageTaskCreationResponse(BaseModel):

comfy_api_nodes/apis/hunyuan3d.py

@@ -77,6 +77,7 @@ class To3DUVTaskRequest(BaseModel):
 
 class To3DPartTaskRequest(BaseModel):
     File: TaskFile3DInput = Field(...)
+    EnableStagedGeneration: bool | None = Field(None)
 
 
 class TextureEditImageInfo(BaseModel):

comfy_api_nodes/apis/sync_so.py

@@ -0,0 +1,49 @@
+from pydantic import BaseModel, Field
+
+
+class SyncInputItem(BaseModel):
+    type: str = Field(..., description="Input kind: 'video', 'image' or 'audio'.")
+    url: str = Field(...)
+
+
+class SyncActiveSpeakerDetection(BaseModel):
+    auto_detect: bool | None = Field(
+        None, description="Detect the active speaker automatically. Video input only; rejected for images."
+    )
+    frame_number: int | None = Field(
+        None, description="Frame used for manual speaker selection. Must be 0 for image inputs."
+    )
+    coordinates: list[int] | None = Field(
+        None, description="Pixel [x, y] of the speaker's face in the frame selected by frame_number."
+    )
+
+
+class SyncGenerationOptions(BaseModel):
+    sync_mode: str | None = Field(
+        None,
+        description="How to resolve an audio/video duration mismatch: "
+        "cut_off, bounce, loop, silence or remap. Ignored for image inputs.",
+    )
+    i2v_prompt: str | None = Field(
+        None, description="Motion prompt for image-to-video generation. Image input only."
+    )
+    active_speaker_detection: SyncActiveSpeakerDetection | None = Field(None)
+
+
+class SyncGenerationRequest(BaseModel):
+    model: str = Field(..., description="Generation model, e.g. 'sync-3'.")
+    input: list[SyncInputItem] = Field(
+        ..., description="Exactly one visual input (video or image) plus one audio input."
+    )
+    options: SyncGenerationOptions | None = Field(None)
+
+
+class SyncGeneration(BaseModel):
+    """Subset of the Generation object returned by POST /v2/generate and GET /v2/generate/{id}."""
+
+    id: str = Field(...)
+    status: str = Field(..., description="PENDING | PROCESSING | COMPLETED | FAILED | REJECTED")
+    outputUrl: str | None = Field(None)
+    outputDuration: float | None = Field(None)
+    error: str | None = Field(None, description="Human-readable failure message.")
+    errorCode: str | None = Field(None, description="Stable machine-readable code from the GET /v2/errors catalog.")

comfy_api_nodes/nodes_bytedance.py

@@ -34,6 +34,7 @@ from comfy_api_nodes.apis.bytedance import (
     SeedanceVirtualLibraryCreateAssetRequest,
     Seedream4Options,
     Seedream4TaskCreationRequest,
+    Seedream5OptimizePromptOptions,
     TaskAudioContent,
     TaskAudioContentUrl,
     TaskCreationResponse,
@@ -875,6 +876,17 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode):
                     tooltip='Whether to add an "AI generated" watermark to the image.',
                     advanced=True,
                 ),
+                IO.Boolean.Input(
+                    "thinking",
+                    default=True,
+                    tooltip=(
+                        "Enable the model's prompt-optimization reasoning ('thinking') for better adherence. "
+                        "Can substantially increase generation time — notably on Seedream 5.0 Pro. "
+                        "Can only be disabled for text-to-image (not when reference images are provided)."
+                    ),
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -920,6 +932,7 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode):
         model: dict,
         seed: int = 0,
         watermark: bool = False,
+        thinking: bool = True,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         model_id = SEEDREAM_MODELS[model["model"]]
@@ -979,6 +992,10 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode):
             raise ValueError(
                 "The maximum number of generated images plus the number of reference images cannot exceed 15."
             )
+        if not thinking and n_input_images > 0:
+            raise ValueError(
+                "'thinking' can only be disabled for text-to-image; enable it when using reference images."
+            )
 
         reference_images_urls: list[str] = []
         if image_tensors:
@@ -992,6 +1009,9 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode):
                 wait_label="Uploading reference images",
             )
 
+        optimize_prompt_options = None
+        if n_input_images == 0:
+            optimize_prompt_options = Seedream5OptimizePromptOptions(thinking="enabled" if thinking else "disabled")
         response = await sync_op(
             cls,
             ApiEndpoint(path=BYTEPLUS_IMAGE_ENDPOINT, method="POST"),
@@ -1005,6 +1025,7 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode):
                 sequential_image_generation=None if is_pro else sequential_image_generation,
                 sequential_image_generation_options=None if is_pro else Seedream4Options(max_images=max_images),
                 watermark=watermark,
+                optimize_prompt_options=optimize_prompt_options,
             ),
         )
         if len(response.data) == 1:

comfy_api_nodes/nodes_gemini.py

@@ -1133,7 +1133,9 @@ class GeminiImage2(IO.ComfyNode):
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         if model == "Nano Banana 2 (Gemini 3.1 Flash Image)":
-            model = "gemini-3.1-flash-image-preview"
+            model = "gemini-3.1-flash-image"
+        elif model == "gemini-3-pro-image-preview":
+            model = "gemini-3-pro-image"
 
         parts: list[GeminiPart] = [GeminiPart(text=prompt)]
         if images is not None:
@@ -1507,7 +1509,7 @@ class GeminiNanoBanana2V2(IO.ComfyNode):
         validate_string(prompt, strip_whitespace=True, min_length=1)
         model_choice = model["model"]
         if model_choice == "Nano Banana 2 (Gemini 3.1 Flash Image)":
-            model_id = "gemini-3.1-flash-image-preview"
+            model_id = "gemini-3.1-flash-image"
         elif model_choice == "Nano Banana 2 Lite":
             model_id = "gemini-3.1-flash-lite-image"
         else:

comfy_api_nodes/nodes_hunyuan3d.py

@@ -642,6 +642,7 @@ class Tencent3DPartNode(IO.ComfyNode):
             response_model=To3DProTaskCreateResponse,
             data=To3DPartTaskRequest(
                 File=TaskFile3DInput(Type=file_format.upper(), Url=model_url),
+                EnableStagedGeneration=True,
             ),
             is_rate_limited=_is_tencent_rate_limited,
         )

comfy_api_nodes/nodes_sync_so.py

@@ -0,0 +1,391 @@
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.sync_so import (
+    SyncActiveSpeakerDetection,
+    SyncGeneration,
+    SyncGenerationOptions,
+    SyncGenerationRequest,
+    SyncInputItem,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    download_url_to_video_output,
+    downscale_image_tensor,
+    downscale_image_tensor_by_max_side,
+    get_image_dimensions,
+    get_number_of_images,
+    poll_op,
+    sync_op,
+    upload_audio_to_comfyapi,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_audio_duration,
+)
+
+
+class SyncLipSyncNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="SyncLipSyncNode",
+            display_name="sync.so Lip Sync",
+            category="partner/video/sync.so",
+            description=(
+                "Re-sync mouth movement in a video to new speech audio using sync.so. "
+                "Handles close-ups, profiles and obstructions automatically while preserving "
+                "the speaker's expression. Cost scales with output duration."
+            ),
+            inputs=[
+                IO.Video.Input(
+                    "video",
+                    tooltip="Footage of the speaker to re-sync. Up to 4K (4096x2160); "
+                    "a constant frame rate of 24/25/30 fps works best.",
+                ),
+                IO.Audio.Input(
+                    "audio",
+                    tooltip="Speech audio to sync the mouth to.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=42,
+                    min=0,
+                    max=2147483647,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "sync-3",
+                            [
+                                IO.Combo.Input(
+                                    "sync_mode",
+                                    options=["bounce", "cut_off", "loop", "silence", "remap"],
+                                    default="bounce",
+                                    tooltip=(
+                                        "How to handle a duration mismatch between video and audio; "
+                                        "this also sets the output length. "
+                                        "bounce: video plays forward then backward until the audio ends "
+                                        "(output = audio length). "
+                                        "loop: video restarts until the audio ends (output = audio length). "
+                                        "remap: video is time-stretched to match the audio (output = audio length). "
+                                        "cut_off: the longer track is trimmed (output = shorter length). "
+                                        "silence: nothing is trimmed; the shorter track is padded "
+                                        "(output = longer length)."
+                                    ),
+                                ),
+                                IO.Combo.Input(
+                                    "speaker_selection",
+                                    options=["default", "auto-detect", "coordinates"],
+                                    default="default",
+                                    tooltip=(
+                                        "Which face to lipsync when several people are visible. "
+                                        "default: let the model decide. "
+                                        "auto-detect: detect and follow the active speaker. "
+                                        "coordinates: target the face at pixel (speaker_x, speaker_y) "
+                                        "in the frame chosen by speaker_frame."
+                                    ),
+                                ),
+                                IO.Int.Input(
+                                    "speaker_frame",
+                                    default=0,
+                                    min=0,
+                                    max=1_000_000,
+                                    advanced=True,
+                                    tooltip="Video frame used to locate the speaker. "
+                                    "Only used when speaker_selection is 'coordinates'.",
+                                ),
+                                IO.Int.Input(
+                                    "speaker_x",
+                                    default=0,
+                                    min=0,
+                                    max=4096,
+                                    advanced=True,
+                                    tooltip="X pixel coordinate of the speaker's face. "
+                                    "Only used when speaker_selection is 'coordinates'.",
+                                ),
+                                IO.Int.Input(
+                                    "speaker_y",
+                                    default=0,
+                                    min=0,
+                                    max=4096,
+                                    advanced=True,
+                                    tooltip="Y pixel coordinate of the speaker's face. "
+                                    "Only used when speaker_selection is 'coordinates'.",
+                                ),
+                            ],
+                        )
+                    ],
+                    tooltip="sync.so generation model.",
+                ),
+            ],
+            outputs=[IO.Video.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(
+                expr="""{"type":"usd","usd":0.19019,"format":{"approximate":true,"suffix":"/second"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        audio: Input.Audio,
+        seed: int,
+        model: dict,
+    ) -> IO.NodeOutput:
+        try:
+            width, height = video.get_dimensions()
+        except Exception:
+            width = height = None
+        if width and height and (max(width, height) > 4096 or width * height > 4096 * 2160):
+            raise ValueError(
+                f"sync.so rejects videos above 4K (4096x2160); got {width}x{height}. Downscale the video first."
+            )
+        validate_audio_duration(audio, max_duration=600)
+
+        if model["speaker_selection"] == "auto-detect":
+            speaker_detection = SyncActiveSpeakerDetection(auto_detect=True)
+        elif model["speaker_selection"] == "coordinates":
+            speaker_detection = SyncActiveSpeakerDetection(
+                frame_number=model["speaker_frame"],
+                coordinates=[model["speaker_x"], model["speaker_y"]],
+            )
+        else:
+            speaker_detection = None
+
+        video_url = await upload_video_to_comfyapi(cls, video, max_duration=600)
+        audio_url = await upload_audio_to_comfyapi(cls, audio)
+
+        generation = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/synclabs/v2/generate", method="POST"),
+            response_model=SyncGeneration,
+            data=SyncGenerationRequest(
+                model=model["model"],
+                input=[
+                    SyncInputItem(type="video", url=video_url),
+                    SyncInputItem(type="audio", url=audio_url),
+                ],
+                options=SyncGenerationOptions(
+                    sync_mode=model["sync_mode"],
+                    active_speaker_detection=speaker_detection,
+                ),
+            ),
+        )
+        generation = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/synclabs/v2/generate/{generation.id}"),
+            response_model=SyncGeneration,
+            status_extractor=lambda g: g.status,
+            completed_statuses=["COMPLETED", "FAILED", "REJECTED"],
+            failed_statuses=[],
+            queued_statuses=["PENDING"],
+            poll_interval=10.0,
+        )
+        if generation.status != "COMPLETED":
+            code = f" [{generation.errorCode}]" if generation.errorCode else ""
+            raise ValueError(
+                f"sync.so generation {generation.status.lower()}{code}: "
+                f"{generation.error or 'no error details provided'}"
+            )
+        if not generation.outputUrl:
+            raise ValueError("sync.so generation completed but no output URL was returned.")
+        return IO.NodeOutput(await download_url_to_video_output(generation.outputUrl))
+
+
+class SyncTalkingImageNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="SyncTalkingImageNode",
+            display_name="sync.so Talking Image",
+            category="partner/video/sync.so",
+            description=(
+                "Animate a still portrait into a talking video driven by speech audio, "
+                "using sync.so's sync-3 model. The output duration matches the audio. "
+                "Cost scales with output duration."
+            ),
+            inputs=[
+                IO.Image.Input(
+                    "image",
+                    tooltip="A single image with a clearly visible face, up to 4K (4096x2160).",
+                ),
+                IO.Audio.Input(
+                    "audio",
+                    tooltip="Speech audio driving the talking video; the output duration matches it. "
+                    "Chain any TTS node here to drive the animation from text.",
+                ),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Optional guidance for how the portrait comes to life, e.g. "
+                    "'make the subject smile and look at the camera'. "
+                    "Leave empty for natural talking motion.",
+                ),
+                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.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "sync-3",
+                            [
+                                IO.Combo.Input(
+                                    "speaker_selection",
+                                    options=["default", "coordinates"],
+                                    default="default",
+                                    tooltip=(
+                                        "Which face to animate when several people are visible. "
+                                        "default: let the model decide. "
+                                        "coordinates: target the face at pixel (speaker_x, speaker_y) "
+                                        "in the image. Auto-detection is not supported for images."
+                                    ),
+                                ),
+                                IO.Int.Input(
+                                    "speaker_x",
+                                    default=0,
+                                    min=0,
+                                    max=4096,
+                                    advanced=True,
+                                    tooltip="X pixel coordinate of the speaker's face. "
+                                    "Only used when speaker_selection is 'coordinates'.",
+                                ),
+                                IO.Int.Input(
+                                    "speaker_y",
+                                    default=0,
+                                    min=0,
+                                    max=4096,
+                                    advanced=True,
+                                    tooltip="Y pixel coordinate of the speaker's face. "
+                                    "Only used when speaker_selection is 'coordinates'.",
+                                ),
+                                IO.Boolean.Input(
+                                    "auto_downscale",
+                                    default=True,
+                                    advanced=True,
+                                    tooltip="Automatically downscale the image if it exceeds the 4K "
+                                    "(4096x2160) input limit; speaker coordinates are scaled to match. "
+                                    "When disabled, an oversized image raises an error instead.",
+                                ),
+                            ],
+                        )
+                    ],
+                    tooltip="sync.so generation model. Image input is exclusive to sync-3.",
+                ),
+            ],
+            outputs=[IO.Video.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(
+                expr="""{"type":"usd","usd":0.19019,"format":{"approximate":true,"suffix":"/second"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        audio: Input.Audio,
+        prompt: str,
+        seed: int,
+        model: dict,
+    ) -> IO.NodeOutput:
+        if get_number_of_images(image) != 1:
+            raise ValueError("Exactly one image is required; got a batch. Pick one frame first.")
+        validate_audio_duration(audio, max_duration=600)
+
+        height, width = get_image_dimensions(image)
+        speaker_x, speaker_y = model["speaker_x"], model["speaker_y"]
+        if max(width, height) > 4096 or width * height > 4096 * 2160:
+            if not model["auto_downscale"]:
+                raise ValueError(
+                    f"sync.so rejects images above 4K (4096x2160); got {width}x{height}. "
+                    "Downscale the image first or enable auto_downscale."
+                )
+            image = downscale_image_tensor(image, total_pixels=4096 * 2160)
+            image = downscale_image_tensor_by_max_side(image, max_side=4096)
+            new_height, new_width = get_image_dimensions(image)
+            # speaker coordinates are given in the original image's pixel space
+            speaker_x = min(new_width - 1, round(speaker_x * new_width / width))
+            speaker_y = min(new_height - 1, round(speaker_y * new_height / height))
+
+        if model["speaker_selection"] == "coordinates":
+            speaker_detection = SyncActiveSpeakerDetection(
+                frame_number=0,  # images have a single frame; auto_detect is rejected by the API
+                coordinates=[speaker_x, speaker_y],
+            )
+        else:
+            speaker_detection = None
+
+        image_url = await upload_image_to_comfyapi(cls, image, mime_type="image/png", total_pixels=None)
+        audio_url = await upload_audio_to_comfyapi(cls, audio)
+
+        generation = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/synclabs/v2/generate", method="POST"),
+            response_model=SyncGeneration,
+            data=SyncGenerationRequest(
+                model=model["model"],
+                input=[
+                    SyncInputItem(type="image", url=image_url),
+                    SyncInputItem(type="audio", url=audio_url),
+                ],
+                options=SyncGenerationOptions(
+                    i2v_prompt=prompt.strip() or None,
+                    active_speaker_detection=speaker_detection,
+                ),
+            ),
+        )
+        generation = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/synclabs/v2/generate/{generation.id}"),
+            response_model=SyncGeneration,
+            status_extractor=lambda g: g.status,
+            completed_statuses=["COMPLETED", "FAILED", "REJECTED"],
+            failed_statuses=[],
+            queued_statuses=["PENDING"],
+            poll_interval=10.0,
+        )
+        if generation.status != "COMPLETED":
+            code = f" [{generation.errorCode}]" if generation.errorCode else ""
+            raise ValueError(
+                f"sync.so generation {generation.status.lower()}{code}: "
+                f"{generation.error or 'no error details provided'}"
+            )
+        if not generation.outputUrl:
+            raise ValueError("sync.so generation completed but no output URL was returned.")
+        return IO.NodeOutput(await download_url_to_video_output(generation.outputUrl))
+
+
+class SyncExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            SyncLipSyncNode,
+            SyncTalkingImageNode,
+        ]
+
+
+async def comfy_entrypoint() -> SyncExtension:
+    return SyncExtension()

comfy_api_nodes/util/_helpers.py

@@ -15,6 +15,7 @@ from comfy.comfy_api_env import normalize_comfy_api_base
 from comfy.deploy_environment import get_deploy_environment
 from comfy.model_management import processing_interrupted
 from comfy_api.latest import IO
+from comfyui_version import __version__ as comfyui_version
 
 from .common_exceptions import ProcessingInterrupted
 
@@ -60,6 +61,7 @@ def get_comfy_api_headers(node_cls: type[IO.ComfyNode]) -> dict[str, str]:
         **get_auth_header(node_cls),
         "Comfy-Env": get_deploy_environment(),
         "Comfy-Usage-Source": get_usage_source(node_cls),
+        "Comfy-Core-Version": comfyui_version,
     }
 
 

comfy_execution/caching.py

@@ -503,6 +503,8 @@ RAM_CACHE_DEFAULT_RAM_USAGE = 0.05
 
 RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER = 1.3
 
+RAM_CACHE_LARGE_INTERMEDIATE = 512 * 1024 ** 2
+
 
 def all_outputs_dynamic(outputs):
     if outputs is None:
@@ -517,7 +519,6 @@ def all_outputs_dynamic(outputs):
 
     return True
 
-
 class RAMPressureCache(LRUCache):
 
     def __init__(self, key_class, enable_providers=False):
@@ -539,9 +540,9 @@ class RAMPressureCache(LRUCache):
         self.timestamps[self.cache_key_set.get_data_key(node_id)] = time.time()
         super().set_local(node_id, value)
 
-    def ram_release(self, target, free_active=False):
+    def ram_release(self, target, free_active=False, min_entry_size=0):
         if psutil.virtual_memory().available >= target:
-            return
+            return 0
 
         clean_list = []
 
@@ -555,8 +556,9 @@ class RAMPressureCache(LRUCache):
             oom_score = RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER ** (self.generation - self.used_generation[key])
 
             ram_usage = RAM_CACHE_DEFAULT_RAM_USAGE
+            oom_ram_usage = ram_usage
             def scan_list_for_ram_usage(outputs):
-                nonlocal ram_usage
+                nonlocal ram_usage, oom_ram_usage
                 if outputs is None:
                     return
                 for output in outputs:
@@ -564,19 +566,26 @@ class RAMPressureCache(LRUCache):
                         scan_list_for_ram_usage(output)
                     elif isinstance(output, torch.Tensor) and output.device.type == 'cpu':
                         ram_usage += output.numel() * output.element_size()
+                        oom_ram_usage += output.numel() * output.element_size()
                     elif isinstance(output, ModelPatcher) and self.used_generation[key] != self.generation:
                         #old ModelPatchers are the first to go
-                        ram_usage = 1e30
+                        oom_ram_usage = 1e30
             scan_list_for_ram_usage(cache_entry.outputs)
 
-            oom_score *= ram_usage
+            if ram_usage < min_entry_size:
+                continue
+
+            oom_score *= oom_ram_usage
             #In the case where we have no information on the node ram usage at all,
             #break OOM score ties on the last touch timestamp (pure LRU)
-            bisect.insort(clean_list, (oom_score, self.timestamps[key], key))
+            bisect.insort(clean_list, (oom_score, self.timestamps[key], key, ram_usage))
 
+        freed = 0
         while psutil.virtual_memory().available < target and clean_list:
-            _, _, key = clean_list.pop()
+            _, _, key, ram_usage = clean_list.pop()
             del self.cache[key]
             self.used_generation.pop(key, None)
             self.timestamps.pop(key, None)
             self.children.pop(key, None)
+            freed += ram_usage
+        return freed

comfy_execution/jobs.py

@@ -56,6 +56,9 @@ PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio', '3d', 'text'})
 # 3D file extensions for preview fallback (no dedicated media_type exists)
 THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb', '.usdz'})
 
+# Text file extensions for preview fallback (the formats SaveText can produce)
+TEXT_EXTENSIONS = frozenset({'.txt', '.md', '.json'})
+
 
 def has_3d_extension(filename: str) -> bool:
     lower = filename.lower()
@@ -143,9 +146,10 @@ def is_previewable(media_type: str, item: dict) -> bool:
     Maintains backwards compatibility with existing logic.
 
     Priority:
-    1. media_type is 'images', 'video', 'audio', or '3d'
+    1. media_type is 'images', 'video', 'audio', '3d', or 'text'
     2. format field starts with 'video/' or 'audio/'
     3. filename has a 3D extension (.obj, .fbx, .gltf, .glb, .usdz)
+    4. filename has a text extension (.txt, .md, .json, ...)
     """
     if media_type in PREVIEWABLE_MEDIA_TYPES:
         return True
@@ -156,10 +160,12 @@ def is_previewable(media_type: str, item: dict) -> bool:
     if fmt and (fmt.startswith('video/') or fmt.startswith('audio/')):
         return True
 
-    # Check for 3D files by extension
+    # Check for 3D and text files by extension
     filename = item.get('filename', '').lower()
     if any(filename.endswith(ext) for ext in THREE_D_EXTENSIONS):
         return True
+    if any(filename.endswith(ext) for ext in TEXT_EXTENSIONS):
+        return True
 
     return False
 
@@ -255,6 +261,10 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]:
     Preview priority (matching frontend):
     1. type="output" with previewable media
     2. Any previewable media
+
+    Text content entries (strings under 'text') are preview-only metadata,
+    matching the frontend's METADATA_KEYS: they can serve as the fallback
+    preview but are not counted as outputs.
     """
     count = 0
     preview_output = None
@@ -275,7 +285,6 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]:
                     if normalized is None:
                         # Not a 3D file string — check for text preview
                         if media_type == 'text':
-                            count += 1
                             if preview_output is None:
                                 if isinstance(item, tuple):
                                     text_value = item[0] if item else ''

comfy_extras/nodes_audio.py

@@ -298,6 +298,7 @@ class PreviewAudio(IO.ComfyNode):
             search_aliases=["play audio"],
             display_name="Preview Audio",
             category="audio",
+            description="Preview the audio without saving it to the ComfyUI output directory.",
             inputs=[
                 IO.Audio.Input("audio"),
             ],

comfy_extras/nodes_bounding_boxes.py

@@ -1,3 +1,5 @@
+import json
+
 import numpy as np
 import torch
 from PIL import Image, ImageDraw, ImageEnhance, ImageFont
@@ -166,6 +168,111 @@ def boxes_to_regions(boxes, width: int, height: int) -> list:
     return regions
 
 
+def normalize_incoming_boxes(bboxes) -> list:
+    if isinstance(bboxes, dict):
+        frame = [bboxes]
+    elif not isinstance(bboxes, list) or not bboxes:
+        frame = []
+    elif isinstance(bboxes[0], dict):
+        frame = bboxes
+    else:
+        frame = bboxes[0] if isinstance(bboxes[0], list) else []
+    boxes = []
+    for box in frame:
+        if not isinstance(box, dict):
+            continue
+        norm = {
+            "x": box.get("x", 0),
+            "y": box.get("y", 0),
+            "width": box.get("width", 0),
+            "height": box.get("height", 0),
+        }
+        meta = box.get("metadata")
+        if isinstance(meta, dict):
+            norm["metadata"] = meta
+        boxes.append(norm)
+    return boxes
+
+
+def _looks_like_element(box: dict) -> bool:
+    bbox = box.get("bbox")
+    return isinstance(bbox, (list, tuple)) and len(bbox) == 4
+
+
+def _looks_like_bbox(box: dict) -> bool:
+    return all(key in box for key in ("x", "y", "width", "height"))
+
+
+def elements_to_boxes(elements: list, width: int, height: int) -> list:
+    boxes = []
+    for element in elements:
+        if not isinstance(element, dict):
+            continue
+        bbox = element.get("bbox")
+        if not (isinstance(bbox, (list, tuple)) and len(bbox) == 4):
+            raise ValueError("bboxes element is missing a valid 'bbox' [ymin, xmin, ymax, xmax]")
+        try:
+            ymin, xmin, ymax, xmax = (float(v) / 1000.0 for v in bbox)
+        except (TypeError, ValueError):
+            raise ValueError("bboxes element 'bbox' must contain four numbers")
+        etype = "text" if element.get("type") == "text" else "obj"
+        boxes.append({
+            "x": round(min(xmin, xmax) * width),
+            "y": round(min(ymin, ymax) * height),
+            "width": round(abs(xmax - xmin) * width),
+            "height": round(abs(ymax - ymin) * height),
+            "metadata": {
+                "type": etype,
+                "text": element.get("text", "") if etype == "text" else "",
+                "desc": element.get("desc", ""),
+                "palette": element.get("color_palette", []) or [],
+            },
+        })
+    return boxes
+
+
+def boxes_from_input(data, width: int, height: int) -> list:
+    if data is None:
+        return []
+    if isinstance(data, str):
+        text = data.strip()
+        if not text:
+            return []
+        try:
+            data = json.loads(text)
+        except (ValueError, TypeError) as exc:
+            raise ValueError(f"bboxes string input is not valid JSON: {exc}") from exc
+    if isinstance(data, dict):
+        if _looks_like_element(data):
+            return elements_to_boxes([data], width, height)
+        if _looks_like_bbox(data):
+            return normalize_incoming_boxes(data)
+        raise ValueError(
+            "bboxes dict must be a bounding box (x, y, width, height) or an element (with a 'bbox')"
+        )
+    if not isinstance(data, list):
+        raise ValueError(
+            "bboxes input must be bounding boxes, elements, or a JSON string, "
+            f"got {type(data).__name__}"
+        )
+    if not data:
+        return []
+    first = data[0]
+    if isinstance(first, list):
+        return normalize_incoming_boxes(data)
+    if isinstance(first, dict):
+        if _looks_like_element(first):
+            return elements_to_boxes(data, width, height)
+        if _looks_like_bbox(first):
+            return normalize_incoming_boxes(data)
+        raise ValueError(
+            "bboxes items must be bounding boxes (x, y, width, height) or elements (with a 'bbox')"
+        )
+    raise ValueError(
+        f"bboxes list must contain bounding boxes or elements, got {type(first).__name__}"
+    )
+
+
 def _norm_bbox(region: dict) -> list[int]:
     def grid(value: float) -> int:
         return max(0, min(1000, round(value * 1000)))
@@ -217,29 +324,48 @@ class CreateBoundingBoxes(io.ComfyNode):
                     optional=True,
                     tooltip="Optional image used as background in the canvas and preview.",
                 ),
+                io.MultiType.Input(
+                    "bboxes",
+                    [io.BoundingBox, io.Array, io.String],
+                    optional=True,
+                    tooltip="Bounding boxes, elements, or a JSON string to initialize the canvas. A new upstream value initializes the canvas; edits made on the canvas take priority and are kept until the upstream value changes again.",
+                ),
                 io.Int.Input("width", default=1024, min=64, max=16384, step=16,
                              tooltip="Width of the canvas and the pixel grid for the bounding boxes."),
                 io.Int.Input("height", default=1024, min=64, max=16384, step=16,
                              tooltip="Height of the canvas and the pixel grid for the bounding boxes."),
                 editor_state,
+                io.BoundingBoxes.Input(
+                    "last_incoming",
+                    optional=True,
+                    tooltip="Internal state managed by the canvas: the upstream bboxes value that last initialized it. Leave empty to re-initialize the canvas from the bboxes input on the next run.",
+                ),
             ],
             outputs=[
                 io.Image.Output(display_name="preview"),
                 io.BoundingBox.Output(display_name="bboxes"),
                 io.Array.Output(display_name="elements"),
             ],
+            is_output_node=True,
             is_experimental=True,
         )
 
     @classmethod
-    def execute(cls, width, height, editor_state=None, background=None) -> io.NodeOutput:
+    def execute(cls, width, height, editor_state=None, last_incoming=None, background=None, bboxes=None) -> io.NodeOutput:
-        regions = boxes_to_regions(editor_state, width, height)
+        incoming = boxes_from_input(bboxes, width, height)
+        applied = last_incoming if isinstance(last_incoming, list) else []
+        upstream_changed = bool(incoming) and incoming != applied
+        source = incoming if upstream_changed else (editor_state or [])
+        regions = boxes_to_regions(source, width, height)
         preview = render_preview(regions, width, height, _bg_from_image(background))
+        ui = {"dims": [width, height]}
+        if incoming:
+            ui["input_bboxes"] = incoming
         return io.NodeOutput(
             preview,
             fractions_to_bbox_frame(regions, width, height),
             build_elements(regions),
-            ui={"dims": [width, height]},
+            ui=ui,
         )
 
 

comfy_extras/nodes_images.py

@@ -844,15 +844,18 @@ class ImageMergeTileList(IO.ComfyNode):
 # Format specifications
 # ---------------------------------------------------------------------------
 
-# Maps (file_format, bit_depth, has_alpha) -> (numpy dtype scale, av pixel format,
+# Maps (file_format, bit_depth, num_channels) -> (quantization scale, numpy dtype,
-# stream pix_fmt). Keeps the encode path declarative instead of branchy.
+# av frame pix_fmt, stream pix_fmt). Keeps the encode path declarative instead of branchy.
 _FORMAT_SPECS = {
-    ("png", "8-bit", False):  {"scale": 255.0,   "dtype": np.uint8,   "frame_fmt": "rgb24",     "stream_fmt": "rgb24"},
+    ("png", "8-bit", 1):  {"scale": 255.0,   "dtype": np.uint8,   "frame_fmt": "gray",      "stream_fmt": "gray"},
-    ("png", "8-bit", True):   {"scale": 255.0,   "dtype": np.uint8,   "frame_fmt": "rgba",      "stream_fmt": "rgba"},
+    ("png", "8-bit", 3):  {"scale": 255.0,   "dtype": np.uint8,   "frame_fmt": "rgb24",     "stream_fmt": "rgb24"},
-    ("png", "16-bit", False): {"scale": 65535.0, "dtype": np.uint16,  "frame_fmt": "rgb48le",   "stream_fmt": "rgb48be"},
+    ("png", "8-bit", 4):  {"scale": 255.0,   "dtype": np.uint8,   "frame_fmt": "rgba",      "stream_fmt": "rgba"},
-    ("png", "16-bit", True):  {"scale": 65535.0, "dtype": np.uint16,  "frame_fmt": "rgba64le",  "stream_fmt": "rgba64be"},
+    ("png", "16-bit", 1): {"scale": 65535.0, "dtype": np.uint16,  "frame_fmt": "gray16le",  "stream_fmt": "gray16be"},
-    ("exr", "32-bit float", False): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrpf32le",  "stream_fmt": "gbrpf32le"},
+    ("png", "16-bit", 3): {"scale": 65535.0, "dtype": np.uint16,  "frame_fmt": "rgb48le",   "stream_fmt": "rgb48be"},
-    ("exr", "32-bit float", True):  {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrapf32le", "stream_fmt": "gbrapf32le"},
+    ("png", "16-bit", 4): {"scale": 65535.0, "dtype": np.uint16,  "frame_fmt": "rgba64le",  "stream_fmt": "rgba64be"},
+    ("exr", "32-bit float", 1): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "grayf32le",  "stream_fmt": "grayf32le"},
+    ("exr", "32-bit float", 3): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrpf32le",  "stream_fmt": "gbrpf32le"},
+    ("exr", "32-bit float", 4): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrapf32le", "stream_fmt": "gbrapf32le"},
 }
 
 
@@ -891,10 +894,11 @@ def hlg_to_linear(t: torch.Tensor) -> torch.Tensor:
         return torch.cat([hlg_to_linear(rgb), alpha], dim=-1)
 
     # Piecewise: sqrt branch below 0.5, log branch above.
-    # Clamp inside the log branch so negative / out-of-range values don't blow up;
+    # Clamp the log branch at the 0.5 branch point (not above it) so the
+    # unselected lane stays finite in exp() without altering selected values;
     # values above 1.0 are allowed and extrapolate naturally.
     low = (t ** 2) / 3.0
-    high = (torch.exp((t.clamp(min=_HLG_C) - _HLG_C) / _HLG_A) + _HLG_B) / 12.0
+    high = (torch.exp((t.clamp(min=0.5) - _HLG_C) / _HLG_A) + _HLG_B) / 12.0
     return torch.where(t <= 0.5, low, high)
 
 
@@ -1087,7 +1091,8 @@ def _encode_image(
     bit_depth: str,
     colorspace: str,
 ) -> bytes:
-    """Encode a single HxWxC tensor to PNG or EXR bytes in memory.
+    """Encode a single HxWxC (or channel-less HxW grayscale) tensor to PNG or
+    EXR bytes in memory. Grayscale is written as single-channel PNG / Y-only EXR.
 
     For EXR the input is interpreted according to `colorspace` and converted
     to scene-linear (EXR's convention) before writing:
@@ -1101,10 +1106,16 @@ def _encode_image(
     For PNG, colorspace selection does not modify pixels — PNG is delivered
     sRGB-encoded and there is no PNG path for wide-gamut HDR in this node.
     """
+    if img_tensor.ndim == 2:
+        img_tensor = img_tensor.unsqueeze(-1)  # Some nodes emit grayscale as (H, W) with no channel dim, mask-style.
     height, width, num_channels = img_tensor.shape
-    has_alpha = num_channels == 4
 
-    spec = _FORMAT_SPECS[(file_format, bit_depth, has_alpha)]
+    spec = _FORMAT_SPECS.get((file_format, bit_depth, num_channels))
+    if spec is None:
+        raise ValueError(
+            f"No {file_format}/{bit_depth} encoder for {num_channels}-channel images: "
+            "supported channel counts are 1 (grayscale), 3 (RGB) and 4 (RGBA)."
+        )
 
     if spec["dtype"] == np.float32:
         # EXR path: preserve full range, no clamp.

comfy_extras/nodes_load_3d.py

@@ -61,14 +61,10 @@ class Load3D(IO.ComfyNode):
 
     @classmethod
     def execute(cls, model_file, image, **kwargs) -> IO.NodeOutput:
-        image_path = folder_paths.get_annotated_filepath(image['image'])
-        mask_path = folder_paths.get_annotated_filepath(image['mask'])
-        normal_path = folder_paths.get_annotated_filepath(image['normal'])
-
         load_image_node = nodes.LoadImage()
-        output_image, ignore_mask = load_image_node.load_image(image=image_path)
+        output_image, ignore_mask = load_image_node.load_image(image=image['image'])
-        ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        ignore_image, output_mask = load_image_node.load_image(image=image['mask'])
-        normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path)
+        normal_image, ignore_mask2 = load_image_node.load_image(image=image['normal'])
 
         video = None
 
@@ -96,6 +92,7 @@ class Preview3D(IO.ComfyNode):
             search_aliases=["view mesh", "3d viewer"],
             display_name="Preview 3D & Animation",
             category="3d",
+            description="Preview a 3D model file without saving it to the ComfyUI output directory.",
             is_experimental=True,
             is_output_node=True,
             inputs=[
@@ -140,6 +137,7 @@ class Preview3DAdvanced(IO.ComfyNode):
             display_name="Preview 3D (Advanced)",
             search_aliases=["preview 3d", "3d viewer", "view mesh", "frame 3d", "3d camera output"],
             category="3d",
+            description="Preview a 3D model file without saving it to the ComfyUI output directory.",
             is_experimental=True,
             is_output_node=True,
             inputs=[
@@ -176,8 +174,9 @@ class Preview3DAdvanced(IO.ComfyNode):
         filename = f"preview3d_advanced_{uuid.uuid4().hex}.{model_3d.format}"
         model_3d.save_to(os.path.join(folder_paths.get_temp_directory(), filename))
 
+        viewport_state = viewport_state if isinstance(viewport_state, dict) else {}
         camera_info_input = kwargs.get("camera_info", None)
-        camera_info = camera_info_input if camera_info_input is not None else viewport_state['camera_info']
+        camera_info = camera_info_input if camera_info_input is not None else viewport_state.get('camera_info')
         model_3d_info_input = kwargs.get("model_3d_info", None)
         model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', [])
         return IO.NodeOutput(
@@ -197,6 +196,7 @@ class PreviewGaussianSplat(IO.ComfyNode):
             node_id="PreviewGaussianSplat",
             display_name="Preview Splat",
             category="3d",
+            description="Preview a gaussian splat 3D file without saving it to the ComfyUI output directory.",
             is_experimental=True,
             is_output_node=True,
             search_aliases=[
@@ -244,8 +244,9 @@ class PreviewGaussianSplat(IO.ComfyNode):
         filename = f"preview_splat_{uuid.uuid4().hex}.{model_3d.format}"
         model_3d.save_to(os.path.join(folder_paths.get_temp_directory(), filename))
 
+        viewport_state = viewport_state if isinstance(viewport_state, dict) else {}
         camera_info_input = kwargs.get("camera_info", None)
-        camera_info = camera_info_input if camera_info_input is not None else viewport_state['camera_info']
+        camera_info = camera_info_input if camera_info_input is not None else viewport_state.get('camera_info')
         model_3d_info_input = kwargs.get("model_3d_info", None)
         model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', [])
         return IO.NodeOutput(
@@ -265,6 +266,7 @@ class PreviewPointCloud(IO.ComfyNode):
             node_id="PreviewPointCloud",
             display_name="Preview Point Cloud",
             category="3d",
+            description="Preview a point cloud 3D file without saving it to the ComfyUI output directory.",
             is_experimental=True,
             is_output_node=True,
             search_aliases=[
@@ -303,8 +305,9 @@ class PreviewPointCloud(IO.ComfyNode):
         filename = f"preview_pointcloud_{uuid.uuid4().hex}.{model_3d.format}"
         model_3d.save_to(os.path.join(folder_paths.get_temp_directory(), filename))
 
+        viewport_state = viewport_state if isinstance(viewport_state, dict) else {}
         camera_info_input = kwargs.get("camera_info", None)
-        camera_info = camera_info_input if camera_info_input is not None else viewport_state['camera_info']
+        camera_info = camera_info_input if camera_info_input is not None else viewport_state.get('camera_info')
         model_3d_info_input = kwargs.get("model_3d_info", None)
         model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', [])
         return IO.NodeOutput(
@@ -375,8 +378,9 @@ class Load3DAdvanced(IO.ComfyNode):
         file_3d = None
         if model_file and model_file != "none":
             file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
+        viewport_state = viewport_state if isinstance(viewport_state, dict) else {}
         model_3d_info = viewport_state.get('model_3d_info', [])
-        return IO.NodeOutput(file_3d, model_3d_info, viewport_state['camera_info'], width, height)
+        return IO.NodeOutput(file_3d, model_3d_info, viewport_state.get('camera_info'), width, height)
 
 
 class Load3DExtension(ComfyExtension):

comfy_extras/nodes_mask.py

@@ -419,17 +419,18 @@ class MaskPreview(IO.ComfyNode):
             search_aliases=["show mask", "view mask", "inspect mask", "debug mask"],
             display_name="Preview Mask",
             category="image/mask",
-            description="Saves the input images to your ComfyUI output directory.",
+            description="Preview the masks without saving them to the ComfyUI output directory.",
             inputs=[
                 IO.Mask.Input("mask"),
             ],
             hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo],
             is_output_node=True,
+            outputs=[IO.Mask.Output(display_name="mask")]
         )
 
     @classmethod
     def execute(cls, mask, filename_prefix="ComfyUI") -> IO.NodeOutput:
-        return IO.NodeOutput(ui=UI.PreviewMask(mask))
+        return IO.NodeOutput(mask, ui=UI.PreviewMask(mask))
 
 
 class MaskExtension(ComfyExtension):

comfy_extras/nodes_preview_any.py

@@ -18,6 +18,7 @@ class PreviewAny():
 
     CATEGORY = "utilities"
     SEARCH_ALIASES = ["show output", "inspect", "debug", "print value", "show text"]
+    DESCRIPTION = "Preview any input value as text."
 
     def main(self, source=None):
         torch.set_printoptions(edgeitems=6)

comfy_extras/nodes_primitive.py

@@ -10,11 +10,10 @@ class String(io.ComfyNode):
         return io.Schema(
             node_id="PrimitiveString",
             search_aliases=["text", "string", "text box", "prompt"],
-            display_name="Text String (DEPRECATED)",
+            display_name="Text",
             category="utilities/primitive",
             inputs=[io.String.Input("value")],
-            outputs=[io.String.Output()],
+            outputs=[io.String.Output()]
-            is_deprecated=True
         )
 
     @classmethod
@@ -28,7 +27,7 @@ class StringMultiline(io.ComfyNode):
         return io.Schema(
             node_id="PrimitiveStringMultiline",
             search_aliases=["text", "string", "text multiline", "string multiline", "text box", "prompt"],
-            display_name="Input Text",
+            display_name="Text (Multiline)",
             category="utilities/primitive",
             essentials_category="Basics",
             inputs=[io.String.Input("value", multiline=True)],

comfy_extras/nodes_save_3d.py

@@ -13,7 +13,7 @@ from typing_extensions import override
 
 import folder_paths
 from comfy.cli_args import args
-from comfy_api.latest import ComfyExtension, IO, Types
+from comfy_api.latest import ComfyExtension, IO, Types, UI
 
 
 def pack_variable_mesh_batch(vertices, faces, colors=None, uvs=None, texture=None, unlit=False):
@@ -406,10 +406,165 @@ class SaveGLB(IO.ComfyNode):
         return IO.NodeOutput(ui={"3d": results})
 
 
+def _save_file3d_to_output(model_3d: Types.File3D, filename_prefix: str) -> str:
+    full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(
+        filename_prefix, folder_paths.get_output_directory()
+    )
+    ext = model_3d.format or "glb"
+    saved_filename = f"{filename}_{counter:05}.{ext}"
+    model_3d.save_to(os.path.join(full_output_folder, saved_filename))
+    return f"{subfolder}/{saved_filename}" if subfolder else saved_filename
+
+
+def execute_save_3d_advanced(model_3d, viewport_state, width, height, filename_prefix, kwargs) -> IO.NodeOutput:
+    model_file = _save_file3d_to_output(model_3d, filename_prefix)
+    viewport_state = viewport_state if isinstance(viewport_state, dict) else {}
+    camera_info_input = kwargs.get("camera_info", None)
+    camera_info = camera_info_input if camera_info_input is not None else viewport_state.get('camera_info')
+    model_3d_info_input = kwargs.get("model_3d_info", None)
+    model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', [])
+    return IO.NodeOutput(
+        model_3d,
+        model_3d_info,
+        camera_info,
+        width,
+        height,
+        ui=UI.PreviewUI3DAdvanced(model_file, camera_info, model_3d_info),
+    )
+
+
+class Save3DAdvanced(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="Save3DAdvanced",
+            display_name="Save 3D (Advanced)",
+            search_aliases=["save 3d", "export 3d model", "save mesh advanced"],
+            category="3d",
+            is_experimental=True,
+            is_output_node=True,
+            inputs=[
+                IO.MultiType.Input(
+                    "model_3d",
+                    types=[
+                        IO.File3DGLB,
+                        IO.File3DGLTF,
+                        IO.File3DFBX,
+                        IO.File3DOBJ,
+                        IO.File3DSTL,
+                        IO.File3DUSDZ,
+                        IO.File3DAny,
+                    ],
+                    tooltip="3D model file from an upstream 3D node.",
+                ),
+                IO.String.Input("filename_prefix", default="3d/ComfyUI"),
+                IO.Load3D.Input("viewport_state"),
+                IO.Load3DModelInfo.Input("model_3d_info", optional=True, advanced=True),
+                IO.Load3DCamera.Input("camera_info", optional=True, advanced=True),
+                IO.Int.Input("width", default=1024, min=1, max=4096, step=1),
+                IO.Int.Input("height", default=1024, min=1, max=4096, step=1),
+            ],
+            outputs=[
+                IO.File3DAny.Output(display_name="model_3d"),
+                IO.Load3DModelInfo.Output(display_name="model_3d_info"),
+                IO.Load3DCamera.Output(display_name="camera_info"),
+                IO.Int.Output(display_name="width"),
+                IO.Int.Output(display_name="height"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_3d: Types.File3D, viewport_state, width: int, height: int, filename_prefix: str, **kwargs) -> IO.NodeOutput:
+        return execute_save_3d_advanced(model_3d, viewport_state, width, height, filename_prefix, kwargs)
+
+
+class SaveGaussianSplat(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="SaveGaussianSplat",
+            display_name="Save Splat",
+            search_aliases=["save splat", "save gaussian splat", "export gaussian", "export splat"],
+            category="3d",
+            is_experimental=True,
+            is_output_node=True,
+            inputs=[
+                IO.MultiType.Input(
+                    "model_3d",
+                    types=[
+                        IO.File3DSplatAny,
+                        IO.File3DPLY,
+                        IO.File3DSPLAT,
+                        IO.File3DSPZ,
+                        IO.File3DKSPLAT,
+                    ],
+                    tooltip="A gaussian splat 3D file.",
+                ),
+                IO.String.Input("filename_prefix", default="3d/ComfyUI"),
+                IO.Load3D.Input("viewport_state"),
+                IO.Load3DModelInfo.Input("model_3d_info", optional=True, advanced=True),
+                IO.Load3DCamera.Input("camera_info", optional=True, advanced=True),
+                IO.Int.Input("width", default=1024, min=1, max=4096, step=1),
+                IO.Int.Input("height", default=1024, min=1, max=4096, step=1),
+            ],
+            outputs=[
+                IO.File3DSplatAny.Output(display_name="model_3d"),
+                IO.Load3DModelInfo.Output(display_name="model_3d_info"),
+                IO.Load3DCamera.Output(display_name="camera_info"),
+                IO.Int.Output(display_name="width"),
+                IO.Int.Output(display_name="height"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_3d: Types.File3D, viewport_state, width: int, height: int, filename_prefix: str, **kwargs) -> IO.NodeOutput:
+        return execute_save_3d_advanced(model_3d, viewport_state, width, height, filename_prefix, kwargs)
+
+
+class SavePointCloud(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="SavePointCloud",
+            display_name="Save Point Cloud",
+            search_aliases=["save point cloud", "save pointcloud", "export point cloud"],
+            category="3d",
+            is_experimental=True,
+            is_output_node=True,
+            inputs=[
+                IO.MultiType.Input(
+                    "model_3d",
+                    types=[
+                        IO.File3DPointCloudAny,
+                        IO.File3DPLY,
+                    ],
+                    tooltip="Point cloud file (.ply)",
+                ),
+                IO.String.Input("filename_prefix", default="3d/ComfyUI"),
+                IO.Load3D.Input("viewport_state"),
+                IO.Load3DModelInfo.Input("model_3d_info", optional=True, advanced=True),
+                IO.Load3DCamera.Input("camera_info", optional=True, advanced=True),
+                IO.Int.Input("width", default=1024, min=1, max=4096, step=1),
+                IO.Int.Input("height", default=1024, min=1, max=4096, step=1),
+            ],
+            outputs=[
+                IO.File3DPointCloudAny.Output(display_name="model_3d"),
+                IO.Load3DModelInfo.Output(display_name="model_3d_info"),
+                IO.Load3DCamera.Output(display_name="camera_info"),
+                IO.Int.Output(display_name="width"),
+                IO.Int.Output(display_name="height"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_3d: Types.File3D, viewport_state, width: int, height: int, filename_prefix: str, **kwargs) -> IO.NodeOutput:
+        return execute_save_3d_advanced(model_3d, viewport_state, width, height, filename_prefix, kwargs)
+
+
 class Save3DExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
-        return [SaveGLB]
+        return [SaveGLB, Save3DAdvanced, SaveGaussianSplat, SavePointCloud]
 
 
 async def comfy_entrypoint() -> Save3DExtension:

comfy_extras/nodes_seedvr.py

@@ -0,0 +1,614 @@
+import logging
+
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io
+import torch
+
+import comfy.model_management
+from comfy.ldm.seedvr.color_fix import (
+    adain_color_transfer,
+    lab_color_transfer,
+    wavelet_color_transfer,
+)
+from comfy.ldm.seedvr.constants import (
+    BYTEDANCE_VAE_SPATIAL_DOWNSAMPLE,
+    SEEDVR2_ADAIN_SCALE_MULTIPLIER,
+    SEEDVR2_CHUNK_GIB_PER_MPX_FRAME,
+    SEEDVR2_CHUNK_RESERVED_GIB,
+    SEEDVR2_CHUNK_SIGMA_GIB,
+    SEEDVR2_CHUNK_SIGMA_K,
+    SEEDVR2_COLOR_MEM_HEADROOM,
+    SEEDVR2_DTYPE_BYTES_FLOOR,
+    SEEDVR2_LAB_SCALE_MULTIPLIER,
+    SEEDVR2_LATENT_CHANNELS,
+    SEEDVR2_OOM_BACKOFF_DIVISOR,
+    SEEDVR2_WAVELET_SCALE_MULTIPLIER,
+)
+
+from torchvision.transforms import functional as TVF
+from torchvision.transforms.functional import InterpolationMode
+
+
+_SEEDVR2_INVALID_MODEL_MSG_PREFIX = "SeedVR2Conditioning: model object does not match expected SeedVR2 structure"
+_ATTR_MISSING = object()
+
+
+def _resolve_seedvr2_diffusion_model(model):
+    inner = getattr(model, "model", _ATTR_MISSING)
+    if inner is _ATTR_MISSING:
+        raise RuntimeError(
+            f"{_SEEDVR2_INVALID_MODEL_MSG_PREFIX}: input has no 'model' attribute "
+            f"(got type {type(model).__name__})."
+        )
+    if inner is None:
+        raise RuntimeError(
+            f"{_SEEDVR2_INVALID_MODEL_MSG_PREFIX}: input.model is None "
+            f"(input type {type(model).__name__})."
+        )
+    diffusion_model = getattr(inner, "diffusion_model", _ATTR_MISSING)
+    if diffusion_model is _ATTR_MISSING:
+        raise RuntimeError(
+            f"{_SEEDVR2_INVALID_MODEL_MSG_PREFIX}: 'model.model' has no "
+            f"'diffusion_model' attribute (got type {type(inner).__name__})."
+        )
+    if diffusion_model is None:
+        raise RuntimeError(
+            f"{_SEEDVR2_INVALID_MODEL_MSG_PREFIX}: 'model.model.diffusion_model' "
+            f"is None (model.model type {type(inner).__name__})."
+        )
+    return diffusion_model
+
+
+def div_pad(image, factor):
+    height_factor, width_factor = factor
+    height, width = image.shape[-2:]
+
+    pad_height = (height_factor - (height % height_factor)) % height_factor
+    pad_width = (width_factor - (width % width_factor)) % width_factor
+
+    if pad_height == 0 and pad_width == 0:
+        return image
+
+    padding = (0, pad_width, 0, pad_height)
+    return torch.nn.functional.pad(image, padding, mode='constant', value=0.0)
+
+def cut_videos(videos):
+    t = videos.size(1)
+    if t < 1:
+        raise ValueError("SeedVR2Preprocess expected at least one frame.")
+    if t == 1:
+        return videos
+    if t <= 4:
+        padding = videos[:, -1:].repeat(1, 4 - t + 1, 1, 1, 1)
+        return torch.cat([videos, padding], dim=1)
+    if (t - 1) % 4 == 0:
+        return videos
+    padding = videos[:, -1:].repeat(1, 4 - ((t - 1) % 4), 1, 1, 1)
+    videos = torch.cat([videos, padding], dim=1)
+    if (videos.size(1) - 1) % 4 != 0:
+        raise ValueError(f"SeedVR2Preprocess failed to pad video length to 4n+1; got {videos.size(1)} frames.")
+    return videos
+
+def _seedvr2_input_shorter_edge(images, node_name):
+    if images.dim() == 4:
+        return min(images.shape[1], images.shape[2])
+    if images.dim() == 5:
+        return min(images.shape[2], images.shape[3])
+    raise ValueError(
+        f"{node_name}: expected 4-D or 5-D IMAGE tensor, "
+        f"got shape {tuple(images.shape)}"
+    )
+
+
+def _seedvr2_pad(images, upscaled_shorter_edge, node_name):
+    if upscaled_shorter_edge < 2:
+        raise ValueError(
+            f"{node_name}: input shorter edge must be at least 2 pixels; "
+            f"got {upscaled_shorter_edge}."
+        )
+    if images.shape[-1] > 3:
+        images = images[..., :3]
+    if images.dim() == 4:
+        # Comfy video components arrive as a 4-D IMAGE frame sequence:
+        # (frames, H, W, C). SeedVR2 consumes that as one video.
+        images = images.unsqueeze(0)
+    elif images.dim() != 5:
+        raise ValueError(
+            f"{node_name}: expected 4-D or 5-D IMAGE tensor, "
+            f"got shape {tuple(images.shape)}"
+        )
+    images = images.permute(0, 1, 4, 2, 3)
+
+    b, t, c, h, w = images.shape
+    images = images.reshape(b * t, c, h, w)
+
+    images = torch.clamp(images, 0.0, 1.0)
+    images = div_pad(images, (16, 16))
+    _, _, new_h, new_w = images.shape
+
+    images = images.reshape(b, t, c, new_h, new_w)
+    images = cut_videos(images)
+    images_bthwc = images.permute(0, 1, 3, 4, 2).contiguous()
+
+    return io.NodeOutput(images_bthwc)
+
+
+class SeedVR2Preprocess(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SeedVR2Preprocess",
+            display_name="Pre-Process SeedVR2 Input",
+            category="image/pre-processors",
+            description="Pad a resized image for SeedVR2 model. Alpha channel is dropped. The node Post-Process SeedVR2 Output re-applies it from the original resized image.",
+            search_aliases=["seedvr2", "upscale", "video upscale", "pad", "preprocess"],
+            inputs=[
+                io.Image.Input("resized_images", tooltip="The resized image to process."),
+            ],
+            outputs=[
+                io.Image.Output("images", tooltip="The padded image for VAE encoding."),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, resized_images):
+        upscaled_shorter_edge = _seedvr2_input_shorter_edge(resized_images, "SeedVR2Preprocess")
+        return _seedvr2_pad(
+            resized_images, upscaled_shorter_edge, "SeedVR2Preprocess",
+        )
+
+
+class SeedVR2PostProcessing(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SeedVR2PostProcessing",
+            display_name="Post-Process SeedVR2 Output",
+            category="image/post-processors",
+            description="Align the generated image with the original resized image and apply color correction.",
+            search_aliases=["seedvr2", "upscale", "color correction", "color match", "postprocess"],
+            inputs=[
+                io.Image.Input("images", tooltip="The generated image to process."),
+                io.Image.Input("original_resized_images", tooltip="The original resized image before pre-processing, used as reference."),
+                io.Combo.Input("color_correction_method", options=["lab", "wavelet", "adain", "none"], default="lab", tooltip="Method to match the generated image colors to the original image. lab: transfer color in CIELAB space, preserving detail (most faithful). wavelet: transfer low-frequency color, keeping upscaled high-frequency detail. adain: match per-channel mean/std (fastest, global tint). none: skip color transfer (geometry alignment only)."),
+            ],
+            outputs=[io.Image.Output(display_name="images", tooltip="The aligned, color-corrected image.")],
+        )
+
+    @classmethod
+    def execute(cls, images, original_resized_images, color_correction_method):
+        alpha_input = None
+        if original_resized_images.shape[-1] == 4:
+            alpha_input = original_resized_images[..., 3:4]
+            original_resized_images = original_resized_images[..., :3]
+        decoded_5d, decoded_was_4d = cls._as_bthwc(images)
+        reference_full, _ = cls._as_bthwc(original_resized_images)
+        decoded_5d = cls._restore_reference_batch_time(decoded_5d, reference_full)
+
+        b = min(decoded_5d.shape[0], reference_full.shape[0])
+        t = min(decoded_5d.shape[1], reference_full.shape[1])
+        reference_h = reference_full.shape[2]
+        reference_w = reference_full.shape[3]
+
+        decoded_5d = decoded_5d[:b, :t, :, :, :]
+        target_h = min(decoded_5d.shape[2], reference_h)
+        target_w = min(decoded_5d.shape[3], reference_w)
+        decoded_5d = decoded_5d[:, :, :target_h, :target_w, :]
+        if color_correction_method in ("lab", "wavelet", "adain"):
+            reference_5d = reference_full[:b, :t, :, :, :]
+            reference_5d = cls._resize_reference(reference_5d, target_h, target_w)
+            output_device = decoded_5d.device
+            decoded_raw = cls._to_seedvr2_raw(decoded_5d)
+            reference_raw = cls._to_seedvr2_raw(reference_5d)
+            decoded_flat = decoded_raw.permute(0, 1, 4, 2, 3).reshape(b * t, decoded_raw.shape[4], target_h, target_w)
+            reference_flat = reference_raw.permute(0, 1, 4, 2, 3).reshape(b * t, reference_raw.shape[4], target_h, target_w)
+            output = cls._color_transfer_chunked(
+                decoded_flat, reference_flat, output_device, color_correction_method,
+            )
+            output = output.reshape(b, t, output.shape[1], output.shape[2], output.shape[3]).permute(0, 1, 3, 4, 2)
+            output = output.add(1.0).div(2.0).clamp(0.0, 1.0)
+        elif color_correction_method == "none":
+            output = decoded_5d
+        else:
+            raise ValueError(f"SeedVR2PostProcessing: unknown color_correction_method {color_correction_method!r}")
+
+        if alpha_input is not None:
+            alpha_5d, _ = cls._as_bthwc(alpha_input)
+            alpha_5d = alpha_5d[:output.shape[0], :output.shape[1], :output.shape[2], :output.shape[3], :]
+            output = torch.cat([output, alpha_5d.to(dtype=output.dtype, device=output.device)], dim=-1)
+        h2 = output.shape[-3] - (output.shape[-3] % 2)
+        w2 = output.shape[-2] - (output.shape[-2] % 2)
+        output = output[:, :, :h2, :w2, :]
+        if decoded_was_4d:
+            output = output.reshape(-1, output.shape[-3], output.shape[-2], output.shape[-1])
+        return io.NodeOutput(output)
+
+    @staticmethod
+    def _as_bthwc(images):
+        if images.ndim == 4:
+            return images.unsqueeze(0), True
+        if images.ndim == 5:
+            return images, False
+        raise ValueError(
+            f"SeedVR2PostProcessing: expected 4-D or 5-D IMAGE tensor, got shape {tuple(images.shape)}"
+        )
+
+    @staticmethod
+    def _restore_reference_batch_time(decoded, reference):
+        if decoded.shape[0] != 1:
+            return decoded
+        ref_b, ref_t = reference.shape[:2]
+        if ref_b < 1 or decoded.shape[1] % ref_b != 0:
+            return decoded
+        decoded_t = decoded.shape[1] // ref_b
+        if decoded_t < ref_t:
+            return decoded
+        return decoded.reshape(ref_b, decoded_t, decoded.shape[2], decoded.shape[3], decoded.shape[4])
+
+    @staticmethod
+    def _to_seedvr2_raw(images):
+        return images.mul(2.0).sub(1.0)
+
+    @staticmethod
+    def _color_transfer_on_vae_device(decoded_flat, reference_flat, output_device, transfer_fn):
+        color_device = comfy.model_management.vae_device()
+        decoded_flat = decoded_flat.to(device=color_device)
+        reference_flat = reference_flat.to(device=color_device)
+        output = transfer_fn(decoded_flat, reference_flat)
+        return output.to(device=output_device)
+
+    @staticmethod
+    def _lab_color_transfer_on_vae_device(decoded_flat, reference_flat, output_device):
+        color_device = comfy.model_management.vae_device()
+        result = None
+        for start in range(decoded_flat.shape[0]):
+            decoded_frame = decoded_flat[start:start + 1].to(device=color_device).clone()
+            reference_frame = reference_flat[start:start + 1].to(device=color_device).clone()
+            output = lab_color_transfer(decoded_frame, reference_frame).to(device=output_device)
+            if result is None:
+                result = torch.empty(
+                    (decoded_flat.shape[0],) + tuple(output.shape[1:]),
+                    device=output_device,
+                    dtype=output.dtype,
+                )
+            result[start:start + 1].copy_(output)
+        if result is None:
+            raise ValueError("SeedVR2PostProcessing: LAB color correction requires at least one frame.")
+        return result
+
+    @classmethod
+    def _color_transfer_chunked(cls, decoded_flat, reference_flat, output_device, color_correction_method):
+        chunk_size = cls._estimate_color_correction_chunk_size(decoded_flat, color_correction_method)
+        while True:
+            try:
+                return cls._run_color_transfer_chunks(
+                    decoded_flat, reference_flat, output_device, color_correction_method, chunk_size,
+                )
+            except Exception as e:
+                comfy.model_management.raise_non_oom(e)
+                if chunk_size <= 1:
+                    raise RuntimeError(
+                        "SeedVR2PostProcessing: color correction OOM at one frame; "
+                        f"color_correction_method={color_correction_method}, shape={tuple(decoded_flat.shape)}."
+                    ) from e
+                chunk_size = max(1, chunk_size // SEEDVR2_OOM_BACKOFF_DIVISOR)
+
+    @classmethod
+    def _run_color_transfer_chunks(cls, decoded_flat, reference_flat, output_device, color_correction_method, chunk_size):
+        result = None
+        for start in range(0, decoded_flat.shape[0], chunk_size):
+            end = min(start + chunk_size, decoded_flat.shape[0])
+            decoded_chunk = decoded_flat[start:end]
+            reference_chunk = reference_flat[start:end]
+            if color_correction_method == "lab":
+                output = cls._lab_color_transfer_on_vae_device(decoded_chunk, reference_chunk, output_device)
+            elif color_correction_method == "wavelet":
+                output = cls._color_transfer_on_vae_device(
+                    decoded_chunk, reference_chunk, output_device, wavelet_color_transfer,
+                )
+            else:
+                output = cls._color_transfer_on_vae_device(
+                    decoded_chunk, reference_chunk, output_device, adain_color_transfer,
+                )
+            if result is None:
+                result = torch.empty(
+                    (decoded_flat.shape[0],) + tuple(output.shape[1:]),
+                    device=output_device,
+                    dtype=output.dtype,
+                )
+            result[start:end].copy_(output)
+        if result is None:
+            raise ValueError("SeedVR2PostProcessing: color correction requires at least one frame.")
+        return result
+
+    @classmethod
+    def _estimate_color_correction_chunk_size(cls, decoded_flat, color_correction_method):
+        multiplier = cls._color_correction_memory_multiplier(color_correction_method)
+        frames = decoded_flat.shape[0]
+        _, channels, height, width = decoded_flat.shape
+        dtype_bytes = max(decoded_flat.element_size(), SEEDVR2_DTYPE_BYTES_FLOOR)
+        bytes_per_frame = height * width * channels * dtype_bytes * multiplier
+        if bytes_per_frame <= 0:
+            return frames
+        color_device = comfy.model_management.vae_device()
+        free_memory = comfy.model_management.get_free_memory(color_device)
+        chunk_size = int((free_memory * SEEDVR2_COLOR_MEM_HEADROOM) // bytes_per_frame)
+        return max(1, min(frames, chunk_size))
+
+    @staticmethod
+    def _color_correction_memory_multiplier(color_correction_method):
+        if color_correction_method == "lab":
+            return SEEDVR2_LAB_SCALE_MULTIPLIER
+        if color_correction_method == "wavelet":
+            return SEEDVR2_WAVELET_SCALE_MULTIPLIER
+        if color_correction_method == "adain":
+            return SEEDVR2_ADAIN_SCALE_MULTIPLIER
+        raise ValueError(f"SeedVR2PostProcessing: unknown color_correction_method {color_correction_method!r}")
+
+    @staticmethod
+    def _resize_reference(reference, height, width):
+        if reference.shape[2] == height and reference.shape[3] == width:
+            return reference
+        b, t = reference.shape[:2]
+        reference_flat = reference.permute(0, 1, 4, 2, 3).reshape(b * t, reference.shape[4], reference.shape[2], reference.shape[3])
+        resized = TVF.resize(
+            reference_flat,
+            size=(height, width),
+            interpolation=InterpolationMode.BICUBIC,
+            antialias=not (isinstance(reference_flat, torch.Tensor) and reference_flat.device.type == "mps"),
+        )
+        return resized.reshape(b, t, resized.shape[1], height, width).permute(0, 1, 3, 4, 2)
+
+
+class SeedVR2Conditioning(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SeedVR2Conditioning",
+            display_name="Apply SeedVR2 Conditioning",
+            category="model/conditioning",
+            description="Build SeedVR2 positive/negative conditioning from a VAE latent.",
+            search_aliases=["seedvr2", "upscale", "conditioning"],
+            inputs=[
+                io.Model.Input("model", tooltip="The SeedVR2 model."),
+                io.Latent.Input("vae_conditioning", display_name="latent"),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive", tooltip="The positive conditioning for sampling."),
+                io.Conditioning.Output(display_name="negative", tooltip="The negative conditioning for sampling."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, vae_conditioning) -> io.NodeOutput:
+
+        vae_conditioning = vae_conditioning["samples"]
+        if vae_conditioning.ndim != 5:
+            raise ValueError(
+                "SeedVR2Conditioning expects a 5-D VAE latent in Comfy "
+                f"channel-first layout; got shape {tuple(vae_conditioning.shape)}."
+            )
+        if vae_conditioning.shape[1] != SEEDVR2_LATENT_CHANNELS:
+            if vae_conditioning.shape[-1] == SEEDVR2_LATENT_CHANNELS:
+                raise ValueError(
+                    "SeedVR2Conditioning expects SeedVR2 VAE latents in Comfy "
+                    f"channel-first layout (B, {SEEDVR2_LATENT_CHANNELS}, T, H, W); "
+                    f"got channel-last shape {tuple(vae_conditioning.shape)}."
+                )
+            raise ValueError(
+                "SeedVR2Conditioning expects SeedVR2 VAE latents with "
+                f"{SEEDVR2_LATENT_CHANNELS} channels; got shape {tuple(vae_conditioning.shape)}."
+            )
+        vae_conditioning = vae_conditioning.movedim(1, -1).contiguous()
+        model = _resolve_seedvr2_diffusion_model(model)
+        pos_cond = model.positive_conditioning
+        neg_cond = model.negative_conditioning
+
+        mask = vae_conditioning.new_ones(vae_conditioning.shape[:-1] + (1,))
+        condition = torch.cat((vae_conditioning, mask), dim=-1)
+        condition = condition.movedim(-1, 1)
+
+        negative = [[neg_cond.unsqueeze(0), {"condition": condition}]]
+        positive = [[pos_cond.unsqueeze(0), {"condition": condition}]]
+
+        return io.NodeOutput(positive, negative)
+
+def _seedvr2_chunk_crossfade_weights(overlap, device, dtype):
+    """Descending previous-chunk weights across the overlap (next chunk gets ``1 - w``): a Hann fade over the middle third, flat shoulders on the outer thirds."""
+    ramp = torch.linspace(0.0, 1.0, steps=overlap, device=device, dtype=dtype)
+    ramp = ((ramp - 1.0 / 3.0) / (1.0 / 3.0)).clamp(0.0, 1.0)
+    return 0.5 + 0.5 * torch.cos(torch.pi * ramp)
+
+
+class SeedVR2TemporalChunk(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SeedVR2TemporalChunk",
+            display_name="Split SeedVR2 Latent",
+            category="model/latent/batch",
+            description="Split a SeedVR2 video latent into overlapping temporal chunks small enough to sample one at a time within VRAM, wiring latents outputs to both Apply SeedVR2 Conditioning and the sampler latent input before recombining with Merge SeedVR2 Latents.",
+            search_aliases=["seedvr2", "split", "chunk", "temporal", "video upscale", "rebatch"],
+            inputs=[
+                io.Latent.Input("latent", tooltip="The VAE-encoded SeedVR2 latent to split."),
+                io.Int.Input("temporal_overlap", default=0, min=0, max=16384,
+                             tooltip="Latent frames shared between adjacent chunks and crossfaded at merge; 0 = no overlap."),
+                io.DynamicCombo.Input("chunking_mode",
+                                      tooltip="manual = use frames_per_chunk exactly; auto = predict the largest chunk that fits free VRAM.",
+                                      options=[
+                                          io.DynamicCombo.Option("auto", []),
+                                          io.DynamicCombo.Option("manual", [
+                                              io.Int.Input("frames_per_chunk", default=21, min=1, max=16384, step=4,
+                                                           tooltip="Pixel frames per temporal chunk (4n+1: 1, 5, 9, 13, ...)."),
+                                          ]),
+                                      ]),
+            ],
+            outputs=[
+                io.Latent.Output(display_name="latents", is_output_list=True,
+                                 tooltip="The temporal chunks in sequence order."),
+                io.Int.Output(display_name="temporal_overlap",
+                              tooltip="The effective latent-frame overlap between adjacent chunks, for Merge SeedVR2 Latents."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, latent, temporal_overlap, chunking_mode) -> io.NodeOutput:
+        samples = latent["samples"]
+        if samples.ndim != 5:
+            raise ValueError(
+                f"SeedVR2TemporalChunk: expected a 5-D video latent (B, C, T, H, W); "
+                f"got shape {tuple(samples.shape)}."
+            )
+        if samples.shape[1] != SEEDVR2_LATENT_CHANNELS:
+            raise ValueError(
+                f"SeedVR2TemporalChunk: expected {SEEDVR2_LATENT_CHANNELS} latent channels; "
+                f"got shape {tuple(samples.shape)}."
+            )
+        if temporal_overlap < 0:
+            raise ValueError(
+                f"SeedVR2TemporalChunk: temporal_overlap must be >= 0; got {temporal_overlap}."
+            )
+        mode = chunking_mode["chunking_mode"]
+        if mode not in ("auto", "manual"):
+            raise ValueError(
+                f"SeedVR2TemporalChunk: chunking_mode must be 'auto' or 'manual'; "
+                f"got {mode!r}."
+            )
+        t_latent = samples.shape[2]
+        t_pixel = 4 * (t_latent - 1) + 1
+
+        if mode == "auto":
+            free_gb = comfy.model_management.get_free_memory(
+                comfy.model_management.get_torch_device()) / (1024 ** 3)
+            mpx_per_frame = (samples.shape[0] * samples.shape[3] * samples.shape[4]) * (BYTEDANCE_VAE_SPATIAL_DOWNSAMPLE ** 2) / 1e6
+            budget_gb = free_gb - SEEDVR2_CHUNK_RESERVED_GIB - SEEDVR2_CHUNK_SIGMA_K * SEEDVR2_CHUNK_SIGMA_GIB
+            chunk_latent_max = max(1, int(budget_gb / (SEEDVR2_CHUNK_GIB_PER_MPX_FRAME * mpx_per_frame)))
+            frames_per_chunk = min(4 * (chunk_latent_max - 1) + 1, t_pixel)
+            logging.info(
+                "SeedVR2TemporalChunk auto: free=%.2fGiB, %.2fMpx -> frames_per_chunk=%d (t_pixel=%d).",
+                free_gb, mpx_per_frame, frames_per_chunk, t_pixel,
+            )
+        else:
+            frames_per_chunk = chunking_mode["frames_per_chunk"]
+            if frames_per_chunk < 1 or (frames_per_chunk - 1) % 4 != 0:
+                raise ValueError(
+                    f"SeedVR2TemporalChunk: frames_per_chunk must be a 4n+1 pixel-frame count "
+                    f"(1, 5, 9, 13, 17, 21, ...); got {frames_per_chunk}."
+                )
+
+        if t_pixel <= frames_per_chunk:
+            return io.NodeOutput([latent], 0)
+
+        chunk_latent = (frames_per_chunk - 1) // 4 + 1
+        temporal_overlap = min(temporal_overlap, chunk_latent - 1)
+        step = chunk_latent - temporal_overlap
+
+        chunks = []
+        for start in range(0, t_latent, step):
+            end = min(start + chunk_latent, t_latent)
+            chunk = latent.copy()
+            chunk["samples"] = samples[:, :, start:end].contiguous()
+            chunks.append(chunk)
+            if end >= t_latent:
+                break
+        return io.NodeOutput(chunks, temporal_overlap)
+
+
+class SeedVR2TemporalMerge(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SeedVR2TemporalMerge",
+            display_name="Merge SeedVR2 Latents",
+            category="model/latent/batch",
+            is_input_list=True,
+            description="Recombine sampled SeedVR2 latent temporal chunks into one latent, crossfading each overlap with a Hann window sized by the temporal_overlap wired from Split SeedVR2 Latent.",
+            search_aliases=["seedvr2", "merge", "temporal", "hann", "crossfade"],
+            inputs=[
+                io.Latent.Input("latents", tooltip="The sampled temporal chunks in sequence order."),
+                io.Int.Input("temporal_overlap", default=0, min=0, max=16384, force_input=True,
+                             tooltip="The temporal_overlap output of Split SeedVR2 Latent. 0 = plain concatenation."),
+            ],
+            outputs=[
+                io.Latent.Output(display_name="latent", tooltip="The recombined full-length latent."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, latents, temporal_overlap) -> io.NodeOutput:
+        temporal_overlap = temporal_overlap[0]
+        if temporal_overlap < 0:
+            raise ValueError(
+                f"SeedVR2TemporalMerge: temporal_overlap must be >= 0; got {temporal_overlap}."
+            )
+        chunks = [entry["samples"] for entry in latents]
+        first = chunks[0]
+        if first.ndim != 5:
+            raise ValueError(
+                f"SeedVR2TemporalMerge: expected 5-D video latents (B, C, T, H, W); "
+                f"chunk 0 has shape {tuple(first.shape)}."
+            )
+        for i, chunk in enumerate(chunks[1:], start=1):
+            if chunk.shape[:2] != first.shape[:2] or chunk.shape[3:] != first.shape[3:]:
+                raise ValueError(
+                    f"SeedVR2TemporalMerge: chunk {i} shape {tuple(chunk.shape)} does not "
+                    f"match chunk 0 shape {tuple(first.shape)} outside the temporal axis."
+                )
+            if i < len(chunks) - 1 and chunk.shape[2] != first.shape[2]:
+                raise ValueError(
+                    f"SeedVR2TemporalMerge: chunk {i} has {chunk.shape[2]} latent frames but "
+                    f"chunk 0 has {first.shape[2]}; only the final chunk may be shorter."
+                )
+
+        out = latents[0].copy()
+        out.pop("noise_mask", None)
+
+        if len(chunks) == 1:
+            out["samples"] = first
+            return io.NodeOutput(out)
+        if temporal_overlap == 0:
+            out["samples"] = torch.cat(chunks, dim=2)
+            return io.NodeOutput(out)
+
+        chunk_latent = first.shape[2]
+        step = chunk_latent - min(temporal_overlap, chunk_latent - 1)
+        t_total = step * (len(chunks) - 1) + chunks[-1].shape[2]
+        b, c, _, h, w = first.shape
+        merged = torch.empty((b, c, t_total, h, w), device=first.device, dtype=first.dtype)
+
+        merged[:, :, :chunk_latent] = first
+        filled = chunk_latent
+        for i, chunk in enumerate(chunks[1:], start=1):
+            start = i * step
+            end = start + chunk.shape[2]
+            # Crossfade width is bounded by the previous fill frontier and by a runt
+            # final chunk shorter than the configured overlap.
+            fade = min(filled - start, chunk.shape[2])
+            if fade > 0:
+                w_prev = _seedvr2_chunk_crossfade_weights(
+                    fade, chunk.device, chunk.dtype).view(1, 1, fade, 1, 1)
+                merged[:, :, start:start + fade] = (
+                    merged[:, :, start:start + fade] * w_prev + chunk[:, :, :fade] * (1.0 - w_prev)
+                )
+                merged[:, :, start + fade:end] = chunk[:, :, fade:]
+            else:
+                merged[:, :, start:end] = chunk
+            filled = end
+
+        out["samples"] = merged
+        return io.NodeOutput(out)
+
+
+class SeedVRExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            SeedVR2Conditioning,
+            SeedVR2Preprocess,
+            SeedVR2PostProcessing,
+            SeedVR2TemporalChunk,
+            SeedVR2TemporalMerge,
+        ]
+
+async def comfy_entrypoint() -> SeedVRExtension:
+    return SeedVRExtension()

comfy_extras/nodes_text.py

@@ -0,0 +1,71 @@
+import os
+import json
+from typing_extensions import override
+from comfy_api.latest import io, ComfyExtension, ui
+import folder_paths
+
+
+class SaveTextNode(io.ComfyNode):
+    """Save text content to .txt, .md, or .json."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SaveText",
+            search_aliases=["save text", "write text", "export text"],
+            display_name="Save Text",
+            category="text",
+            description="Save text content to a file in the output directory.",
+            inputs=[
+                io.String.Input("text", force_input=True),
+                io.String.Input("filename_prefix", default="ComfyUI"),
+                io.Combo.Input("format", options=["txt", "md", "json"], default="txt"),
+            ],
+            outputs=[io.String.Output(display_name="text")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, text, filename_prefix, format):
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(
+            filename_prefix,
+            folder_paths.get_output_directory(),
+            1,
+            1,
+        )
+
+        file = f"{filename}_{counter:05}.{format}"
+        filepath = os.path.join(full_output_folder, file)
+
+        if format == "json":
+            # tries to pretty print otherwise saves normally
+            try:
+                data = json.loads(text)
+                with open(filepath, "w", encoding="utf-8") as f:
+                    json.dump(data, f, indent=2, ensure_ascii=False)
+            except json.JSONDecodeError:
+                with open(filepath, "w", encoding="utf-8") as f:
+                    f.write(text)
+        else:
+            with open(filepath, "w", encoding="utf-8") as f:
+                f.write(text)
+
+        return io.NodeOutput(
+            text,
+            ui={
+                "text": (text,),
+                "files": [
+                    ui.SavedResult(file, subfolder, io.FolderType.output)
+                ]
+            }
+        )
+
+class TextExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            SaveTextNode
+        ]
+
+async def comfy_entrypoint() -> TextExtension:
+    return TextExtension()

comfy_extras/nodes_video.py

@@ -81,7 +81,7 @@ class SaveVideo(io.ComfyNode):
             display_name="Save Video",
             category="video",
             essentials_category="Basics",
-            description="Saves the input images to your ComfyUI output directory.",
+            description="Saves the input videos to your ComfyUI output directory.",
             inputs=[
                 io.Video.Input("video", tooltip="The video to save."),
                 io.String.Input("filename_prefix", default="video/ComfyUI", tooltip="The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."),

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.27.0"
+__version__ = "0.28.0"

execution.py

@@ -29,6 +29,7 @@ from comfy_execution.caching import (
     HierarchicalCache,
     LRUCache,
     RAMPressureCache,
+    RAM_CACHE_LARGE_INTERMEDIATE,
 )
 from comfy_execution.graph import (
     DynamicPrompt,
@@ -425,12 +426,12 @@ def _is_intermediate_output(dynprompt, node_id):
 
 
 def _send_cached_ui(server, node_id, display_node_id, cached, prompt_id, ui_outputs):
+    if cached.ui is not None:
+        ui_outputs[node_id] = cached.ui
     if server.client_id is None:
         return
     cached_ui = cached.ui or {}
     server.send_sync("executed", { "node": node_id, "display_node": display_node_id, "output": cached_ui.get("output", None), "prompt_id": prompt_id }, server.client_id)
-    if cached.ui is not None:
-        ui_outputs[node_id] = cached.ui
 
 async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_outputs):
     unique_id = current_item
@@ -794,12 +795,16 @@ class PromptExecutor:
                     if self.cache_type == CacheType.RAM_PRESSURE:
                         ram_release_callback(ram_inactive_headroom)
                         ram_shortfall = ram_headroom - psutil.virtual_memory().available
-                        freed = comfy.model_management.free_pins(ram_shortfall + 512 * (1024 ** 2))
+                        if ram_shortfall > 0:
-                        if freed < ram_shortfall:
+                            freed = ram_release_callback(ram_headroom, free_active=True, min_entry_size=RAM_CACHE_LARGE_INTERMEDIATE)
-                            if freed > 64 * (1024 ** 2):
+                            ram_shortfall -= freed
-                                # AIMDO MEM_DECOMMIT can outrun psutil.available catching up.
+                        if comfy.model_management.should_free_pins_for_ram_pressure(ram_shortfall):
-                                time.sleep(0.05)
+                            freed = comfy.model_management.free_pins(ram_shortfall + 512 * (1024 ** 2))
-                            ram_release_callback(ram_headroom, free_active=True)
+                            if freed < ram_shortfall:
+                                if freed > 64 * (1024 ** 2):
+                                    # AIMDO MEM_DECOMMIT can outrun psutil.available catching up.
+                                    time.sleep(0.05)
+                                ram_release_callback(ram_headroom, free_active=True)
                 else:
                     # Only execute when the while-loop ends without break
                     # Send cached UI for intermediate output nodes that weren't executed

nodes.py

@@ -1709,6 +1709,7 @@ class PreviewImage(SaveImage):
         self.compress_level = 1
 
     SEARCH_ALIASES = ["preview", "preview image", "show image", "view image", "display image", "image viewer"]
+    DESCRIPTION = "Preview the images without saving them to the ComfyUI output directory."
 
     @classmethod
     def INPUT_TYPES(s):
@@ -2458,6 +2459,7 @@ async def init_builtin_extra_nodes():
         "nodes_camera_trajectory.py",
         "nodes_edit_model.py",
         "nodes_tcfg.py",
+        "nodes_seedvr.py",
         "nodes_context_windows.py",
         "nodes_qwen.py",
         "nodes_boogu.py",
@@ -2503,6 +2505,7 @@ async def init_builtin_extra_nodes():
         "nodes_triposplat.py",
         "nodes_depth_anything_3.py",
         "nodes_seed.py",
+        "nodes_text.py",
     ]
 
     import_failed = []

openapi.yaml

@@ -7,18 +7,18 @@ components:
                     description: Timestamp when the asset was created
                     format: date-time
                     type: string
+                display_name:
+                    description: Display name of the asset. Mirrors name for backwards compatibility.
+                    nullable: true
+                    type: string
+                file_path:
+                    description: Relative path in global-namespace-root form (e.g. "models/checkpoints/flux.safetensors")
+                    nullable: true
+                    type: string
                 hash:
                     description: Blake3 hash of the asset content.
                     pattern: ^blake3:[a-f0-9]{64}$
                     type: string
-                loader_path:
-                    description: The value a loader consumes to load this asset. Null when no loader can resolve the file.
-                    nullable: true
-                    type: string
-                display_name:
-                    description: Human-facing label for the asset. Not unique.
-                    nullable: true
-                    type: string
                 id:
                     description: Unique identifier for the asset
                     format: uuid
@@ -144,6 +144,14 @@ components:
         AssetUpdated:
             description: Response returned when an existing asset is successfully updated.
             properties:
+                display_name:
+                    description: Display name of the asset. Mirrors name for backwards compatibility.
+                    nullable: true
+                    type: string
+                file_path:
+                    description: Relative path in global-namespace-root form (e.g. "models/checkpoints/flux.safetensors")
+                    nullable: true
+                    type: string
                 hash:
                     description: Blake3 hash of the asset content.
                     pattern: ^blake3:[a-f0-9]{64}$
@@ -1636,7 +1644,7 @@ paths:
                                     format: uuid
                                     type: string
                                 tags:
-                                    description: JSON-encoded array of tag strings. For new byte uploads, include exactly one destination role (`input`, `output`, or `models`); `models` uploads also require exactly one `model_type:<folder_name>` tag. Extra tags are stored as labels and do not create path components.
+                                    description: JSON-encoded array of freeform tag strings, e.g. '["models","checkpoint"]'. Common types include "models", "input", "output", and "temp", but any tag can be used in any order.
                                     type: string
                                 user_metadata:
                                     description: Custom JSON metadata as a string
@@ -1821,7 +1829,7 @@ paths:
                     content:
                         application/json:
                             schema:
-                                $ref: '#/components/schemas/Asset'
+                                $ref: '#/components/schemas/AssetUpdated'
                     description: Asset updated successfully
                 "400":
                     content:
@@ -2462,9 +2470,6 @@ paths:
                                     supports_preview_metadata:
                                         description: Whether the server supports preview metadata
                                         type: boolean
-                                    supports_model_type_tags:
-                                        description: Whether the server supports namespaced model type asset tags
-                                        type: boolean
                                 type: object
                     description: Success
                     headers:
@@ -3292,6 +3297,12 @@ paths:
                             schema:
                                 $ref: '#/components/schemas/ErrorResponse'
                     description: Invalid request parameters
+                "401":
+                    content:
+                        application/json:
+                            schema:
+                                $ref: '#/components/schemas/ErrorResponse'
+                    description: Unauthorized - Authentication required
                 "500":
                     content:
                         application/json:

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.27.0"
+version = "0.28.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,6 +1,6 @@
-comfyui-frontend-package==1.45.20
+comfyui-frontend-package==1.45.21
-comfyui-workflow-templates==0.11.6
+comfyui-workflow-templates==0.11.9
-comfyui-embedded-docs==0.5.7
+comfyui-embedded-docs==0.5.8
 torch
 torchsde
 torchvision
@@ -22,7 +22,7 @@ alembic
 SQLAlchemy>=2.0.0
 filelock
 av>=16.0.0
-comfy-kitchen==0.2.16
+comfy-kitchen==0.2.20
 comfy-aimdo==0.4.10
 requests
 simpleeval>=1.0.0

tests-unit/app_test/model_manager_test.py

@@ -24,6 +24,28 @@ def app(model_manager):
     app.add_routes(routes)
     return app
 
+async def test_get_model_folders_includes_registered_extensions(aiohttp_client, app, tmp_path):
+    """Folders expose their registered extension set verbatim; an empty list
+    means match-all (filter_files_extensions semantics)."""
+    with patch('folder_paths.folder_names_and_paths', {
+        'test_checkpoints': ([str(tmp_path)], {'.safetensors', '.ckpt'}),
+        'test_configs': ([str(tmp_path)], ['.yaml']),
+        'test_match_all': ([str(tmp_path)], set()),
+        'configs': ([str(tmp_path)], ['.yaml']),
+    }):
+        client = await aiohttp_client(app)
+        response = await client.get('/experiment/models')
+
+        assert response.status == 200
+        folders = {f['name']: f for f in await response.json()}
+
+        assert 'configs' not in folders  # blocklisted
+        assert folders['test_checkpoints']['folders'] == [str(tmp_path)]
+        assert folders['test_checkpoints']['extensions'] == ['.ckpt', '.safetensors']
+        assert folders['test_configs']['extensions'] == ['.yaml']
+        # Match-all registrations are exposed honestly, not substituted.
+        assert folders['test_match_all']['extensions'] == []
+
 async def test_get_model_preview_safetensors(aiohttp_client, app, tmp_path):
     img = Image.new('RGB', (100, 100), 'white')
     img_byte_arr = BytesIO()

tests-unit/comfy_extras_test/test_seedvr2_conditioning.py

@@ -0,0 +1,186 @@
+"""SeedVR2 conditioning node regression tests."""
+
+import importlib
+import sys
+from unittest.mock import MagicMock
+
+import pytest
+import torch
+import torch.nn as nn
+
+from comfy.cli_args import args as cli_args
+from comfy.ldm.seedvr.constants import SEEDVR2_LATENT_CHANNELS
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+
+_SENTINEL = object()
+_TARGETS = (
+    ("comfy.model_management", "comfy"),
+    ("comfy_extras.nodes_seedvr", "comfy_extras"),
+)
+
+
+def _import_nodes_seedvr_isolated():
+    """Import comfy_extras.nodes_seedvr with comfy.model_management mocked."""
+    priors = []
+    for mod_name, parent_name in _TARGETS:
+        prior_mod = sys.modules.get(mod_name, _SENTINEL)
+        parent = sys.modules.get(parent_name)
+        attr = mod_name.split(".")[-1]
+        prior_attr = (
+            getattr(parent, attr, _SENTINEL) if parent is not None else _SENTINEL
+        )
+        priors.append((mod_name, parent_name, attr, prior_mod, prior_attr))
+
+    mock_mm = MagicMock()
+    for fn in (
+        "xformers_enabled", "xformers_enabled_vae",
+        "pytorch_attention_enabled", "pytorch_attention_enabled_vae",
+        "sage_attention_enabled", "flash_attention_enabled",
+        "is_intel_xpu",
+    ):
+        getattr(mock_mm, fn).return_value = False
+    tv = torch.version.__version__.split(".")
+    mock_mm.torch_version_numeric = (int(tv[0]), int(tv[1]))
+    mock_mm.WINDOWS = False
+    sys.modules["comfy.model_management"] = mock_mm
+    if sys.modules.get("comfy") is None:
+        importlib.import_module("comfy")
+    comfy_pkg = sys.modules.get("comfy")
+    if comfy_pkg is not None:
+        setattr(comfy_pkg, "model_management", mock_mm)
+    nodes_seedvr = sys.modules.get("comfy_extras.nodes_seedvr") or (
+        importlib.import_module("comfy_extras.nodes_seedvr")
+    )
+
+    def _restore():
+        for mod_name, parent_name, attr, prior_mod, prior_attr in priors:
+            if prior_mod is _SENTINEL:
+                sys.modules.pop(mod_name, None)
+            else:
+                sys.modules[mod_name] = prior_mod
+            parent = sys.modules.get(parent_name)
+            if parent is None:
+                continue
+            if prior_attr is _SENTINEL:
+                if hasattr(parent, attr):
+                    delattr(parent, attr)
+            else:
+                setattr(parent, attr, prior_attr)
+
+    return nodes_seedvr, _restore
+
+
+class _Rope(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.freqs = nn.Parameter(torch.zeros(4))
+
+
+class _Block(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.rope = _Rope()
+
+
+class _DiffusionModel(nn.Module):
+    def __init__(self, n_blocks=3, conditioning_dtype=torch.float32):
+        super().__init__()
+        self.blocks = nn.ModuleList([_Block() for _ in range(n_blocks)])
+        self.register_buffer("positive_conditioning", torch.ones((2, 4), dtype=conditioning_dtype))
+        self.register_buffer("negative_conditioning", torch.zeros((3, 4), dtype=conditioning_dtype))
+
+
+class _ModelInner:
+    def __init__(self, diffusion_model):
+        self.diffusion_model = diffusion_model
+
+
+class _ModelPatcher:
+    def __init__(self, diffusion_model):
+        self.model = _ModelInner(diffusion_model)
+
+
+def test_seedvr2_conditioning_schema_exposes_conditioning_outputs():
+    nodes_seedvr, restore = _import_nodes_seedvr_isolated()
+    try:
+        schema = nodes_seedvr.SeedVR2Conditioning.define_schema()
+        assert [input_item.id for input_item in schema.inputs] == [
+            "model",
+            "vae_conditioning",
+        ]
+        assert schema.inputs[1].display_name == "latent"
+        assert [output.display_name for output in schema.outputs] == [
+            "positive",
+            "negative",
+        ]
+    finally:
+        restore()
+
+
+def test_seedvr2_conditioning_rejects_wrong_latent_channels():
+    nodes_seedvr, restore = _import_nodes_seedvr_isolated()
+    try:
+        patcher = _ModelPatcher(_DiffusionModel())
+        vae_conditioning = {"samples": torch.zeros(1, 8, 2, 2, 2)}
+
+        with pytest.raises(ValueError, match=f"{SEEDVR2_LATENT_CHANNELS} channels"):
+            nodes_seedvr.SeedVR2Conditioning.execute(patcher, vae_conditioning)
+    finally:
+        restore()
+
+
+def test_seedvr2_conditioning_returns_conditioning_deterministically():
+    nodes_seedvr, restore = _import_nodes_seedvr_isolated()
+    try:
+        diffusion_model = _DiffusionModel()
+        patcher = _ModelPatcher(diffusion_model)
+        samples = torch.arange(
+            1,
+            1 + SEEDVR2_LATENT_CHANNELS * 3 * 2 * 2,
+            dtype=torch.float32,
+        ).reshape(1, SEEDVR2_LATENT_CHANNELS, 3, 2, 2)
+        vae_conditioning = {"samples": samples}
+
+        first_positive, first_negative = (
+            nodes_seedvr.SeedVR2Conditioning.execute(
+                patcher,
+                vae_conditioning,
+            )
+        )
+        second_positive, second_negative = (
+            nodes_seedvr.SeedVR2Conditioning.execute(
+                patcher,
+                vae_conditioning,
+            )
+        )
+
+        channel_last = samples.movedim(1, -1).contiguous()
+        expected_condition = torch.cat(
+            [
+                channel_last,
+                torch.ones((*channel_last.shape[:-1], 1)),
+            ],
+            dim=-1,
+        ).movedim(-1, 1)
+
+        assert torch.equal(
+            first_positive[0][1]["condition"],
+            expected_condition,
+        )
+        assert torch.equal(
+            second_positive[0][1]["condition"],
+            expected_condition,
+        )
+        assert torch.equal(
+            first_negative[0][1]["condition"],
+            expected_condition,
+        )
+        assert torch.equal(
+            second_negative[0][1]["condition"],
+            expected_condition,
+        )
+    finally:
+        restore()

tests-unit/comfy_extras_test/test_seedvr2_nodes.py

@@ -0,0 +1,55 @@
+import importlib
+import inspect
+import sys
+from unittest.mock import MagicMock, patch
+
+import torch
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+
+def test_seedvr_node_signature_matches_schema():
+    mock_mm = MagicMock()
+    mock_mm.xformers_enabled.return_value = False
+    mock_mm.xformers_enabled_vae.return_value = False
+    mock_mm.sage_attention_enabled.return_value = False
+    mock_mm.flash_attention_enabled.return_value = False
+
+    sentinel = object()
+    prior_cpu = cli_args.cpu
+    cli_args.cpu = True
+    prior_module = sys.modules.get("comfy_extras.nodes_seedvr", sentinel)
+    comfy_pkg = sys.modules.get("comfy")
+    prior_mm_attr = getattr(comfy_pkg, "model_management", sentinel) if comfy_pkg else sentinel
+
+    with patch.dict(sys.modules, {"comfy.model_management": mock_mm}):
+        if comfy_pkg is not None:
+            setattr(comfy_pkg, "model_management", mock_mm)
+        sys.modules.pop("comfy_extras.nodes_seedvr", None)
+        try:
+            nodes_seedvr = importlib.import_module("comfy_extras.nodes_seedvr")
+            for node_cls in (nodes_seedvr.SeedVR2Preprocess, nodes_seedvr.SeedVR2PostProcessing, nodes_seedvr.SeedVR2Conditioning):
+                schema_ids = [i.id for i in node_cls.define_schema().inputs]
+                exec_params = [
+                    p for p in inspect.signature(node_cls.execute).parameters.keys()
+                    if p != "cls"
+                ]
+                assert schema_ids == exec_params, (
+                    f"{node_cls.__name__} schema/execute drift: "
+                    f"schema_ids={schema_ids}, exec_params={exec_params}"
+                )
+        finally:
+            cli_args.cpu = prior_cpu
+            if prior_module is sentinel:
+                sys.modules.pop("comfy_extras.nodes_seedvr", None)
+            else:
+                sys.modules["comfy_extras.nodes_seedvr"] = prior_module
+            if comfy_pkg is not None:
+                if prior_mm_attr is sentinel:
+                    if hasattr(comfy_pkg, "model_management"):
+                        delattr(comfy_pkg, "model_management")
+                else:
+                    setattr(comfy_pkg, "model_management", prior_mm_attr)

tests-unit/comfy_extras_test/test_seedvr2_post_processing.py

@@ -0,0 +1,51 @@
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+from comfy_extras import nodes_seedvr  # noqa: E402
+
+
+def _schema_ids(items):
+    return [item.id for item in items]
+
+
+def test_seedvr2_post_processing_schema():
+    schema = nodes_seedvr.SeedVR2PostProcessing.define_schema()
+
+    assert _schema_ids(schema.inputs) == ["images", "original_resized_images", "color_correction_method"]
+    assert schema.inputs[2].options == ["lab", "wavelet", "adain", "none"]
+    assert schema.inputs[2].default == "lab"
+    assert schema.outputs[0].get_io_type() == "IMAGE"
+
+
+def test_seedvr2_post_processing_oom_error_uses_color_correction_method(monkeypatch):
+    decoded = torch.full((1, 3, 4, 4), 0.25)
+    reference = torch.full((1, 3, 4, 4), 0.75)
+
+    def _lab(content, style):
+        raise torch.cuda.OutOfMemoryError("CUDA out of memory")
+
+    monkeypatch.setattr(nodes_seedvr.comfy.model_management, "vae_device", lambda: torch.device("cpu"))
+    monkeypatch.setattr(nodes_seedvr.comfy.model_management, "get_free_memory", lambda device: 1_000_000)
+
+    with patch.object(nodes_seedvr, "lab_color_transfer", _lab):
+        with pytest.raises(RuntimeError) as excinfo:
+            nodes_seedvr.SeedVR2PostProcessing._color_transfer_chunked(
+                decoded, reference, torch.device("cpu"), "lab",
+            )
+    assert "color_correction_method=lab" in str(excinfo.value)
+    assert " method=lab" not in str(excinfo.value)
+
+
+def test_seedvr2_post_processing_unknown_color_correction_method_raises():
+    decoded = torch.zeros(1, 2, 4, 4, 3)
+    original = torch.zeros(1, 2, 4, 4, 3)
+    with pytest.raises(ValueError) as excinfo:
+        nodes_seedvr.SeedVR2PostProcessing.execute(decoded, original, "bogus")
+    assert "color_correction_method" in str(excinfo.value)

tests-unit/comfy_extras_test/test_seedvr2_temporal_chunk.py

@@ -0,0 +1,77 @@
+"""SeedVR2 temporal chunk/merge node regression tests."""
+
+import pytest
+import torch
+
+from comfy.cli_args import args as cli_args
+from comfy.ldm.seedvr.constants import (
+    BYTEDANCE_VAE_SPATIAL_DOWNSAMPLE,
+    SEEDVR2_CHUNK_GIB_PER_MPX_FRAME,
+    SEEDVR2_CHUNK_RESERVED_GIB,
+    SEEDVR2_CHUNK_SIGMA_GIB,
+    SEEDVR2_CHUNK_SIGMA_K,
+    SEEDVR2_LATENT_CHANNELS,
+)
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+import comfy.model_management  # noqa: E402
+from comfy_extras.nodes_seedvr import SeedVR2TemporalChunk, SeedVR2TemporalMerge, _seedvr2_chunk_crossfade_weights  # noqa: E402
+
+def _latent(t_latent, h=8, w=8, b=1):
+    g = torch.Generator().manual_seed(7)
+    return {"samples": torch.randn(b, SEEDVR2_LATENT_CHANNELS, t_latent, h, w, generator=g)}
+
+def _split(latent, frames_per_chunk, temporal_overlap, chunking_mode="manual"):
+    combo = {"chunking_mode": chunking_mode}
+    if chunking_mode != "auto":
+        combo["frames_per_chunk"] = frames_per_chunk
+    return SeedVR2TemporalChunk.execute(latent, temporal_overlap, combo).args
+
+def _merge(chunks, temporal_overlap):
+    return SeedVR2TemporalMerge.execute(chunks, [temporal_overlap]).args[0]
+
+def test_chunk_temporal_windows_and_validation():
+    with pytest.raises(ValueError, match="4n\\+1"):
+        _split(_latent(9), 20, 0)
+    with pytest.raises(ValueError, match="5-D"):
+        _split({"samples": torch.zeros(1, SEEDVR2_LATENT_CHANNELS * 9, 8, 8)}, 21, 0)
+    with pytest.raises(ValueError, match="chunking_mode"):
+        _split(_latent(13), 21, 0, "adaptive")
+    latent = _latent(13)
+    chunks, overlap = _split(latent, 21, 2)  # chunk_latent=6, step=4 -> [0:6], [4:10], [8:13]
+    assert overlap == 2 and [c["samples"].shape[2] for c in chunks] == [6, 6, 5]
+    assert all(torch.equal(c["samples"], latent["samples"][:, :, s:e]) for c, (s, e) in zip(chunks, [(0, 6), (4, 10), (8, 13)]))
+    assert len(_split(_latent(13), 21, 999)[0]) == 8  # overlap clamps to chunk_latent-1 -> step=1
+    assert (r := _split(_latent(5), 21, 3)) and len(r[0]) == 1 and r[1] == 0  # t_pixel <= 21: passthrough
+
+def test_chunk_auto_mode_applies_vram_law(monkeypatch):
+    mpx_per_frame = (32 * 32) * (BYTEDANCE_VAE_SPATIAL_DOWNSAMPLE ** 2) / 1e6
+    free_gb = (
+        SEEDVR2_CHUNK_RESERVED_GIB
+        + SEEDVR2_CHUNK_SIGMA_K * SEEDVR2_CHUNK_SIGMA_GIB
+        + 5.1 * SEEDVR2_CHUNK_GIB_PER_MPX_FRAME * mpx_per_frame
+    )
+    monkeypatch.setattr(comfy.model_management, "get_free_memory", lambda dev=None: free_gb * (1024 ** 3))
+    assert [c["samples"].shape[2] for c in _split(_latent(13, h=32, w=32), 1, 0, "auto")[0]] == [5, 5, 3]
+    assert _split(_latent(13, h=32, w=32, b=2), 1, 0, "auto")[0][0]["samples"].shape[2] == 2  # batch halves the chunk
+
+def test_merge_crossfade_and_reassembly():
+    latent = _latent(13)
+    latent["noise_mask"] = torch.rand(1, 1, 13, 8, 8)
+    latent["batch_index"] = [0]
+    merged = _merge(_split(latent, 21, 0)[0], 0)
+    assert torch.equal(merged["samples"], latent["samples"])
+    assert "noise_mask" not in merged and merged["batch_index"] == [0]
+    assert torch.allclose(_merge(_split(latent, 21, 3)[0], 3)["samples"], latent["samples"], atol=1e-6)
+    w = _seedvr2_chunk_crossfade_weights(3, merged["samples"].device, merged["samples"].dtype)
+    assert w[0] == 1.0 and w[-1] == 0.0 and torch.all(w[:-1] >= w[1:])
+    ones, zeros = {"samples": torch.ones(1, SEEDVR2_LATENT_CHANNELS, 6, 8, 8)}, {"samples": torch.zeros(1, SEEDVR2_LATENT_CHANNELS, 6, 8, 8)}
+    fused = _merge([ones, zeros], 3)["samples"]  # overlap equals w: prev fades out, next fades in
+    assert torch.equal(fused[:, :, 3:6], w.view(1, 1, 3, 1, 1).expand(1, SEEDVR2_LATENT_CHANNELS, 3, 8, 8))
+    assert torch.equal(fused[:, :, :3], ones["samples"][:, :, :3]) and torch.equal(fused[:, :, 6:], zeros["samples"][:, :, :3])
+    short = _split(latent, 21, 2)[0]
+    short[0]["samples"] = short[0]["samples"][:, :, :4]
+    with pytest.raises(ValueError, match="only the final chunk may be shorter"):
+        _merge(short, 2)

tests-unit/comfy_quant/test_mixed_precision.py

@@ -15,7 +15,7 @@ if not has_gpu():
     args.cpu = True
 
 from comfy import ops
-from comfy.quant_ops import QuantizedTensor
+from comfy.quant_ops import QUANT_ALGOS, QuantizedTensor
 import comfy.utils
 
 
@@ -283,7 +283,59 @@ class TestMixedPrecisionOps(unittest.TestCase):
         saved = model.state_dict()
         saved_conf = json.loads(saved["layer.comfy_quant"].numpy().tobytes())
         self.assertTrue(saved_conf["convrot"])
+
+    def test_convrot_w4a4_loads_into_params(self):
+        """ConvRot W4A4 checkpoints must load as the dedicated kitchen layout."""
+        if "convrot_w4a4" not in QUANT_ALGOS:
+            self.skipTest("comfy_kitchen does not provide ConvRot W4A4")
+
+        torch.manual_seed(456)
+        layer_quant_config = {
+            "layer": {
+                "format": "convrot_w4a4",
+                "convrot_groupsize": 256,
+                "linear_dtype": "int8",
+            }
+        }
+        weight = torch.randn(16, 256, dtype=torch.bfloat16)
+        bias = torch.randn(16, dtype=torch.bfloat16)
+        q_weight = QuantizedTensor.from_float(
+            weight,
+            "TensorCoreConvRotW4A4Layout",
+            convrot_groupsize=256,
+            quant_group_size=64,
+        )
+        state_dict = {
+            "layer.weight": q_weight._qdata,
+            "layer.bias": bias,
+            "layer.weight_scale": q_weight._params.scale,
+        }
+
+        state_dict, _ = comfy.utils.convert_old_quants(
+            state_dict,
+            metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})},
+        )
+        model = torch.nn.Module()
+        model.layer = ops.mixed_precision_ops({}).Linear(256, 16, device="cpu", dtype=torch.bfloat16)
+        model.load_state_dict(state_dict, strict=False)
+
+        self.assertIsInstance(model.layer.weight, QuantizedTensor)
+        self.assertEqual(model.layer.weight._layout_cls, "TensorCoreConvRotW4A4Layout")
+        self.assertEqual(model.layer.weight._params.convrot_groupsize, 256)
+        self.assertEqual(model.layer.weight._params.quant_group_size, 64)
+        self.assertEqual(model.layer.weight._params.linear_dtype, "int8")
+
+        input_tensor = torch.randn(4, 256, dtype=torch.bfloat16)
+        loaded_out = model.layer(input_tensor)
+        ref_out = torch.nn.functional.linear(input_tensor, q_weight, bias)
+        self.assertTrue(torch.equal(loaded_out, ref_out))
+
+        saved = model.state_dict()
+        saved_conf = json.loads(saved["layer.comfy_quant"].numpy().tobytes())
+        self.assertEqual(saved_conf["format"], "convrot_w4a4")
         self.assertEqual(saved_conf["convrot_groupsize"], 256)
+        self.assertEqual(saved_conf["linear_dtype"], "int8")
+        self.assertNotIn("quant_group_size", saved_conf)
 
 if __name__ == "__main__":
     unittest.main()

tests-unit/comfy_test/model_detection_test.py

@@ -2,7 +2,7 @@ from collections import defaultdict
 
 import torch
 
-from comfy.model_detection import detect_unet_config, model_config_from_unet_config
+from comfy.model_detection import detect_unet_config, model_config_from_unet, model_config_from_unet_config
 import comfy.supported_models
 
 
@@ -73,6 +73,49 @@ def _make_flux_schnell_comfyui_sd():
     return sd
 
 
+def _make_seedvr2_7b_separate_mm_sd():
+    return {
+        "blocks.35.mlp.vid.proj_out.weight": torch.empty(3072, 1),
+        "positive_conditioning": torch.empty(58, 5120),
+        "negative_conditioning": torch.empty(64, 5120),
+    }
+
+
+def _make_seedvr2_7b_shared_mm_sd():
+    return {
+        "blocks.35.mlp.all.proj_in_gate.weight": torch.empty(1, 1),
+        "positive_conditioning": torch.empty(58, 5120),
+        "negative_conditioning": torch.empty(64, 5120),
+    }
+
+
+def _make_seedvr2_3b_shared_mm_sd():
+    return {
+        "blocks.31.mlp.all.proj_in_gate.weight": torch.empty(1, 1),
+        "positive_conditioning": torch.empty(58, 5120),
+        "negative_conditioning": torch.empty(64, 5120),
+    }
+
+
+def _make_pid_v1_5_sd(latent_proj_channels=16):
+    sd = {
+        "pixel_embedder.proj.weight": torch.empty(16, 3, device="meta"),
+        "lq_proj.latent_proj.0.weight": torch.empty(1024, latent_proj_channels, 3, 3, device="meta"),
+        "lq_proj.pit_head.weight": torch.empty(1536, 1024, device="meta"),
+        "lq_proj.gate_modules.0.content_proj.weight": torch.empty(1, 3072, device="meta"),
+        "pixel_blocks.0.attn.q_norm.weight": torch.empty(72, device="meta"),
+        "pixel_blocks.0.adaLN_modulation.0.weight": torch.empty(24576, 1536, device="meta"),
+        "pixel_blocks.0.adaLN_modulation.0.bias": torch.empty(24576, device="meta"),
+    }
+    for i in range(7):
+        sd[f"lq_proj.gate_modules.{i}.log_alpha"] = torch.empty((), device="meta")
+    return sd
+
+
+def _add_model_diffusion_prefix(sd):
+    return {f"model.diffusion_model.{k}": v for k, v in sd.items()}
+
+
 class TestModelDetection:
     """Verify that first-match model detection selects the correct model
     based on list ordering and unet_config specificity."""
@@ -125,6 +168,96 @@ class TestModelDetection:
         assert model_config is not None
         assert type(model_config).__name__ == "FluxSchnell"
 
+    def test_seedvr2_7b_separate_mm_detection_config(self):
+        sd = _make_seedvr2_7b_separate_mm_sd()
+        unet_config = detect_unet_config(sd, "")
+
+        assert unet_config is not None
+        assert unet_config["image_model"] == "seedvr2"
+        assert unet_config["vid_dim"] == 3072
+        assert unet_config["heads"] == 24
+        assert unet_config["num_layers"] == 36
+        assert unet_config["mm_layers"] == 36
+        assert unet_config["mlp_type"] == "normal"
+        assert unet_config["rope_type"] == "rope3d"
+        assert unet_config["rope_dim"] == 64
+
+    def test_seedvr2_7b_shared_mm_detection_config(self):
+        sd = _make_seedvr2_7b_shared_mm_sd()
+        unet_config = detect_unet_config(sd, "")
+
+        assert unet_config is not None
+        assert unet_config["image_model"] == "seedvr2"
+        assert unet_config["vid_dim"] == 3072
+        assert unet_config["heads"] == 24
+        assert unet_config["num_layers"] == 36
+        assert unet_config["mm_layers"] == 10
+        assert unet_config["mlp_type"] == "swiglu"
+        assert unet_config["rope_type"] == "rope3d"
+        assert unet_config["rope_dim"] == 64
+
+    def test_seedvr2_3b_shared_mm_detection_config(self):
+        sd = _make_seedvr2_3b_shared_mm_sd()
+        unet_config = detect_unet_config(sd, "")
+
+        assert unet_config is not None
+        assert unet_config["image_model"] == "seedvr2"
+        assert unet_config["vid_dim"] == 2560
+        assert unet_config["heads"] == 20
+        assert unet_config["num_layers"] == 32
+        assert unet_config["mlp_type"] == "swiglu"
+
+    def test_seedvr2_model_match_requires_conditioning_tensors(self):
+        sd = _make_seedvr2_7b_shared_mm_sd()
+        unet_config = detect_unet_config(sd, "")
+
+        assert type(model_config_from_unet_config(unet_config, sd)).__name__ == "SeedVR2"
+
+        del sd["positive_conditioning"]
+        assert model_config_from_unet_config(unet_config, sd) is None
+
+    def test_seedvr2_model_match_accepts_full_checkpoint_prefix(self):
+        sd = _add_model_diffusion_prefix(_make_seedvr2_7b_shared_mm_sd())
+
+        assert type(model_config_from_unet(sd, "model.diffusion_model.")).__name__ == "SeedVR2"
+
+    def test_pid_v1_5_detection(self):
+        sd = _make_pid_v1_5_sd()
+        unet_config = detect_unet_config(sd, "")
+
+        assert unet_config == {
+            "image_model": "pid",
+            "lq_latent_channels": 16,
+            "lq_hidden_dim": 1024,
+            "latent_spatial_down_factor": 8,
+            "lq_interval": 2,
+            "lq_latent_unpatchify_factor": 1,
+            "lq_conv_padding_mode": "replicate",
+            "lq_gate_per_token": True,
+            "pit_lq_inject": True,
+            "rope_ref_h": 2048,
+            "rope_ref_w": 2048,
+        }
+        assert type(model_config_from_unet_config(unet_config, sd)).__name__ == "PiD"
+
+    def test_pid_v1_5_flux2_detection(self):
+        unet_config = detect_unet_config(_make_pid_v1_5_sd(latent_proj_channels=32), "")
+
+        assert unet_config["lq_latent_channels"] == 128
+        assert unet_config["latent_spatial_down_factor"] == 16
+        assert unet_config["lq_latent_unpatchify_factor"] == 2
+
+    def test_pid_v1_5_pixel_adaln_conversion(self):
+        sd = _make_pid_v1_5_sd()
+        model_config = model_config_from_unet_config(detect_unet_config(sd, ""), sd)
+        processed = model_config.process_unet_state_dict(sd)
+
+        assert processed["pixel_blocks.0.attn.q_norm.weight"].shape == (72,)
+        assert processed["pixel_blocks.0.adaLN_modulation_msa.weight"].shape == (12288, 1536)
+        assert processed["pixel_blocks.0.adaLN_modulation_mlp.weight"].shape == (12288, 1536)
+        assert processed["pixel_blocks.0.adaLN_modulation_msa.bias"].shape == (12288,)
+        assert processed["pixel_blocks.0.adaLN_modulation_mlp.bias"].shape == (12288,)
+
     def test_unet_config_and_required_keys_combination_is_unique(self):
         """Each model in the registry must have a unique combination of
         ``unet_config`` and ``required_keys``. If two models share the same

tests-unit/comfy_test/seedvr_vae_forward_test.py

@@ -0,0 +1,74 @@
+"""Regression tests for the SeedVR2 VAE forward return contract."""
+
+import pytest
+import torch
+import torch.nn as nn
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+from comfy.ldm.seedvr.vae import SEEDVR2_LATENT_CHANNELS, VideoAutoencoderKL  # noqa: E402
+
+
+_LATENT_SHAPE = (1, SEEDVR2_LATENT_CHANNELS, 2, 2, 2)
+_DECODED_SHAPE = (1, 3, 5, 16, 16)
+_INPUT_ENCODE_SHAPE = (1, 3, 5, 16, 16)
+_INPUT_DECODE_SHAPE = _LATENT_SHAPE
+
+
+class _StubVAE(VideoAutoencoderKL):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self._encode_out = torch.zeros(*_LATENT_SHAPE)
+        self._decode_out = torch.zeros(*_DECODED_SHAPE)
+
+    def encode(self, x, return_dict=True):
+        return self._encode_out
+
+    def decode_(self, z, return_dict=True):
+        return self._decode_out
+
+
+def test_forward_encode_returns_tensor():
+    vae = _StubVAE()
+    x = torch.zeros(*_INPUT_ENCODE_SHAPE)
+    result = vae.forward(x, mode="encode")
+    assert type(result) is torch.Tensor
+    assert result.shape == torch.Size(_LATENT_SHAPE)
+
+
+def test_forward_decode_returns_tensor():
+    vae = _StubVAE()
+    z = torch.zeros(*_INPUT_DECODE_SHAPE)
+    result = vae.forward(z, mode="decode")
+    assert type(result) is torch.Tensor
+    assert result.shape == torch.Size(_DECODED_SHAPE)
+
+
+class _TupleReturningStubVAE(VideoAutoencoderKL):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self._encode_tensor = torch.zeros(*_LATENT_SHAPE)
+        self._decode_tensor = torch.zeros(*_DECODED_SHAPE)
+
+    def encode(self, x, return_dict=True):
+        return (self._encode_tensor,)
+
+    def decode_(self, z, return_dict=True):
+        return (self._decode_tensor,)
+
+
+def test_forward_all_unwraps_one_tuple_at_each_step():
+    vae = _TupleReturningStubVAE()
+    x = torch.zeros(*_INPUT_ENCODE_SHAPE)
+    result = vae.forward(x, mode="all")
+    assert type(result) is torch.Tensor
+    assert result.shape == torch.Size(_DECODED_SHAPE)
+
+
+def test_forward_rejects_unknown_mode():
+    vae = _StubVAE()
+    with pytest.raises(ValueError, match="Unknown SeedVR2 VAE forward mode"):
+        vae.forward(torch.zeros(*_INPUT_ENCODE_SHAPE), mode="bogus")

tests-unit/comfy_test/test_seedvr2_dtype.py

@@ -0,0 +1,79 @@
+import torch
+import torch.nn as nn
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+import comfy.sd
+import comfy.supported_models
+import comfy.ldm.seedvr.model as seedvr_model
+import comfy.ldm.seedvr.vae as seedvr_vae
+
+
+def test_seedvr2_fp16_manual_cast_only_for_bf16_device(monkeypatch):
+    bf16_device = object()
+    fp16_device = object()
+
+    monkeypatch.setattr(
+        comfy.supported_models.comfy.model_management,
+        "should_use_bf16",
+        lambda device=None: device is bf16_device,
+    )
+
+    bf16_config = comfy.supported_models.SeedVR2({"image_model": "seedvr2"})
+    bf16_config.set_inference_dtype(torch.float16, None, device=bf16_device)
+    assert bf16_config.manual_cast_dtype is torch.bfloat16
+
+    fp16_config = comfy.supported_models.SeedVR2({"image_model": "seedvr2"})
+    fp16_config.set_inference_dtype(torch.float16, None, device=fp16_device)
+    assert fp16_config.manual_cast_dtype is None
+
+
+def test_seedvr2_text_conditioning_accepts_cfg1_single_branch():
+    context = torch.arange(6, dtype=torch.float32).reshape(1, 3, 2)
+
+    txt, txt_shape = seedvr_model.NaDiT._resolve_text_conditioning(object(), context, [0])
+
+    torch.testing.assert_close(txt, context.squeeze(0))
+    torch.testing.assert_close(txt_shape, torch.tensor([[3]], device=context.device))
+
+
+def test_seedvr2_vae_decode_memory_covers_full_frame_lab_transfer():
+    wrapper = seedvr_vae.VideoAutoencoderKLWrapper.__new__(seedvr_vae.VideoAutoencoderKLWrapper)
+    latent_channels = seedvr_vae.SEEDVR2_LATENT_CHANNELS
+    estimate = wrapper.comfy_memory_used_decode((1, latent_channels, 26, 120, 160))
+    old_estimate = latent_channels * 120 * 160 * (4 * 8 * 8) * 2
+
+    assert estimate == 101 * 960 * 1280 * 160
+    assert estimate > 15 * 1024 ** 3
+    assert estimate > old_estimate * 100
+
+
+def test_seedvr2_vae_encode_preserves_compute_dtype(monkeypatch):
+    wrapper = seedvr_vae.VideoAutoencoderKLWrapper.__new__(seedvr_vae.VideoAutoencoderKLWrapper)
+    nn.Module.__init__(wrapper)
+    wrapper._dummy = nn.Parameter(torch.empty(1, dtype=torch.float16))
+    input_dtype = None
+
+    def encode(self, x):
+        nonlocal input_dtype
+        input_dtype = x.dtype
+        return x
+
+    monkeypatch.setattr(seedvr_vae.VideoAutoencoderKL, "encode", encode)
+
+    x = torch.zeros((1, 3, 1, 8, 8), dtype=torch.float32)
+    wrapper._encode_with_raw_latent(x)
+
+    assert input_dtype == torch.float32
+
+
+def test_seedvr2_vae_ops_cast_weights_to_compute_dtype():
+    attention = seedvr_vae.Attention(query_dim=4, heads=1, dim_head=4).to(torch.float16)
+    hidden_states = torch.zeros((1, 2, 4), dtype=torch.float32)
+
+    output = attention(hidden_states)
+
+    assert output.dtype == torch.float32

tests-unit/comfy_test/test_seedvr2_internals.py

@@ -0,0 +1,169 @@
+"""SeedVR2 internals regression tests."""
+
+from __future__ import annotations
+
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from comfy.cli_args import args
+
+if not torch.cuda.is_available():
+    args.cpu = True
+
+import comfy.ldm.seedvr.model as seedvr_model  # noqa: E402
+import comfy.ldm.seedvr.vae as vae_mod  # noqa: E402
+import comfy.ldm.modules.attention as attention  # noqa: E402
+import comfy.ops as comfy_ops  # noqa: E402
+from comfy.ldm.seedvr.vae import (  # noqa: E402
+    causal_norm_wrapper,
+    set_norm_limit,
+)
+from comfy.ldm.seedvr.attention import var_attention_optimized_split  # noqa: E402
+
+
+_NUM_CHANNELS = 8
+_NUM_GROUPS = 4
+_TENSOR_SHAPE = (1, 8, 2, 4, 4)
+
+_GROUPNORM_SUBCLASSES = [
+    pytest.param(comfy_ops.disable_weight_init.GroupNorm, id="disable_weight_init"),
+    pytest.param(comfy_ops.manual_cast.GroupNorm, id="manual_cast"),
+]
+
+
+@pytest.mark.parametrize("groupnorm_cls", _GROUPNORM_SUBCLASSES)
+def test_seedvr_groupnorm_low_limit_uses_chunked_groupnorm_path(groupnorm_cls):
+    real_group_norm = vae_mod.F.group_norm
+    set_norm_limit(1e-9)
+    try:
+        gn = groupnorm_cls(num_channels=_NUM_CHANNELS, num_groups=_NUM_GROUPS)
+        gn.eval()
+
+        forward_hook_calls = []
+
+        def _hook(module, inputs, output):
+            forward_hook_calls.append(tuple(inputs[0].shape))
+
+        spy_calls = []
+
+        def _group_norm_spy(input_tensor, num_groups_arg, *args, **kwargs):
+            spy_calls.append({"num_groups": int(num_groups_arg)})
+            return real_group_norm(input_tensor, num_groups_arg, *args, **kwargs)
+
+        handle = gn.register_forward_hook(_hook)
+        try:
+            with patch.object(vae_mod.F, "group_norm", side_effect=_group_norm_spy):
+                out_tensor = causal_norm_wrapper(gn, torch.randn(*_TENSOR_SHAPE))
+        finally:
+            handle.remove()
+
+        full_calls = len(forward_hook_calls)
+        chunked_calls = sum(1 for entry in spy_calls if entry["num_groups"] < _NUM_GROUPS)
+
+        assert tuple(int(s) for s in out_tensor.shape) == _TENSOR_SHAPE
+        assert full_calls == 0, (
+            f"low-limit GroupNorm gate must NOT take the full-forward path; got full_calls={full_calls}"
+        )
+        assert chunked_calls > 0, (
+            f"low-limit GroupNorm gate must take the chunked path; got chunked_calls={chunked_calls}"
+        )
+    finally:
+        set_norm_limit(None)
+
+
+def test_seedvr2_7b_swin_attention_forward_uses_optimized_var_attention(monkeypatch):
+    dim = 8
+    heads = 2
+    head_dim = 4
+    attn = seedvr_model.NaSwinAttention(
+        vid_dim=dim,
+        txt_dim=dim,
+        heads=heads,
+        head_dim=head_dim,
+        qk_bias=False,
+        qk_norm=comfy_ops.disable_weight_init.RMSNorm,
+        qk_norm_eps=1e-6,
+        rope_type=None,
+        rope_dim=head_dim,
+        shared_weights=False,
+        window=(2, 1, 1),
+        window_method="720pwin_by_size_bysize",
+        version=True,
+        device="cpu",
+        dtype=torch.float32,
+        operations=comfy_ops.disable_weight_init,
+    )
+    generator = torch.Generator(device="cpu").manual_seed(11)
+    vid = torch.randn(8, dim, generator=generator)
+    txt = torch.randn(3, dim, generator=generator)
+    vid_shape = torch.tensor([[2, 2, 2]], dtype=torch.long)
+    txt_shape = torch.tensor([[3]], dtype=torch.long)
+    calls = []
+
+    def fake_optimized_var_attention(**kwargs):
+        calls.append(kwargs)
+        return kwargs["q"]
+
+    monkeypatch.setattr(seedvr_model, "optimized_var_attention", fake_optimized_var_attention)
+
+    vid_out, txt_out = attn(vid, txt, vid_shape, txt_shape, seedvr_model.Cache(disable=True))
+
+    assert tuple(vid_out.shape) == (8, dim)
+    assert tuple(txt_out.shape) == (3, dim)
+    assert len(calls) == 1
+    call = calls[0]
+    assert tuple(call["q"].shape) == (14, heads, head_dim)
+    assert tuple(call["k"].shape) == (14, heads, head_dim)
+    assert tuple(call["v"].shape) == (14, heads, head_dim)
+    assert call["heads"] == heads
+    assert call["skip_reshape"] is True
+    assert call["skip_output_reshape"] is True
+    assert call["cu_seqlens_q"] == [0, 7, 14]
+    assert call["cu_seqlens_k"] == [0, 7, 14]
+
+
+def test_var_attention_optimized_split_calls_dense_backend_per_window(monkeypatch):
+    heads = 2
+    head_dim = 3
+    q = torch.arange(30, dtype=torch.float32).reshape(5, heads, head_dim)
+    k = q + 100
+    v = q + 200
+    cu = [0, 2, 5]
+    calls = []
+
+    def fake_optimized_attention(q_arg, k_arg, v_arg, heads_arg, **kwargs):
+        calls.append(
+            {
+                "q_shape": tuple(q_arg.shape),
+                "k_shape": tuple(k_arg.shape),
+                "v_shape": tuple(v_arg.shape),
+                "heads": heads_arg,
+                "kwargs": kwargs,
+            }
+        )
+        return q_arg + v_arg
+
+    monkeypatch.setattr(attention, "optimized_attention", fake_optimized_attention)
+
+    out = var_attention_optimized_split(
+        q,
+        k,
+        v,
+        heads,
+        cu,
+        cu,
+        skip_reshape=True,
+        skip_output_reshape=True,
+    )
+
+    assert tuple(out.shape) == (5, heads, head_dim)
+    assert len(calls) == 2
+    assert calls[0]["q_shape"] == (1, heads, 2, head_dim)
+    assert calls[1]["q_shape"] == (1, heads, 3, head_dim)
+    assert all(call["heads"] == heads for call in calls)
+    assert all(call["kwargs"]["skip_reshape"] is True for call in calls)
+    assert all(call["kwargs"]["skip_output_reshape"] is True for call in calls)
+    torch.testing.assert_close(out, q + v, rtol=0, atol=0)
+

tests-unit/comfy_test/test_seedvr2_model.py

@@ -0,0 +1,320 @@
+"""SeedVR2 model, latent-format, and VAE graph regression tests."""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+import torch
+from torch import nn
+
+from comfy.cli_args import args
+
+if not torch.cuda.is_available():
+    args.cpu = True
+
+import comfy  # noqa: E402
+import comfy.latent_formats  # noqa: E402
+import comfy.ldm.seedvr.model as seedvr_model  # noqa: E402
+import comfy.ldm.seedvr.vae as seedvr_vae_mod  # noqa: E402
+import comfy.model_management  # noqa: E402
+import comfy.ops as comfy_ops  # noqa: E402
+import comfy.sample  # noqa: E402
+import comfy.sd as sd_mod  # noqa: E402
+import nodes as nodes_mod  # noqa: E402
+from comfy.ldm.seedvr.model import NaDiT  # noqa: E402
+
+
+_LATENT_CHANNELS = seedvr_vae_mod.SEEDVR2_LATENT_CHANNELS
+
+
+def _make_standin(positive_conditioning):
+    class _StandIn(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.register_buffer(
+                "positive_conditioning", positive_conditioning
+            )
+
+        _resolve_text_conditioning = NaDiT._resolve_text_conditioning
+
+    return _StandIn()
+
+
+class _StubModule(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+
+def _capture_last_layer_flags(monkeypatch, vid_dim: int, txt_in_dim: int) -> list[bool]:
+    flags = []
+
+    class _Block(_StubModule):
+        def __init__(self, *args, **kwargs):
+            flags.append(kwargs["is_last_layer"])
+            super().__init__()
+
+    monkeypatch.setattr(seedvr_model, "NaPatchIn", _StubModule)
+    monkeypatch.setattr(seedvr_model, "NaPatchOut", _StubModule)
+    monkeypatch.setattr(seedvr_model, "TimeEmbedding", _StubModule)
+    monkeypatch.setattr(seedvr_model, "NaMMSRTransformerBlock", _Block)
+
+    seedvr_model.NaDiT(
+        norm_eps=1e-5,
+        num_layers=4,
+        mlp_type="normal",
+        vid_dim=vid_dim,
+        txt_in_dim=txt_in_dim,
+        heads=24,
+        mm_layers=3,
+        operations=comfy_ops.disable_weight_init,
+    )
+
+    return flags
+
+
+class _Model:
+    def __init__(self, latent_format):
+        self._latent_format = latent_format
+
+    def get_model_object(self, name):
+        assert name == "latent_format"
+        return self._latent_format
+
+
+class _Patcher:
+    def get_free_memory(self, device):
+        return 1024 * 1024 * 1024
+
+
+class _EncodeWrapper(seedvr_vae_mod.VideoAutoencoderKLWrapper):
+    def __init__(self, encoded):
+        nn.Module.__init__(self)
+        self.encoded = encoded
+        self.spatial_downsample_factor = 8
+        self.temporal_downsample_factor = 4
+        self.seen = []
+
+    def encode(self, x):
+        self.seen.append(tuple(x.shape))
+        return self.encoded.to(device=x.device, dtype=x.dtype)
+
+
+class _DecodeWrapper(seedvr_vae_mod.VideoAutoencoderKLWrapper):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self.spatial_downsample_factor = 8
+        self.temporal_downsample_factor = 4
+        self.calls = []
+
+    def decode(self, z, seedvr2_tiling=None):
+        self.calls.append({"shape": tuple(z.shape), "seedvr2_tiling": seedvr2_tiling})
+        if z.ndim == 4:
+            b, tc, h, w = z.shape
+            t = tc // _LATENT_CHANNELS
+        else:
+            b, _, t, h, w = z.shape
+        return torch.zeros(b, 3, t, h * 8, w * 8, dtype=z.dtype, device=z.device)
+
+
+def test_seedvr2_wrapper_public_encode_returns_tensor(monkeypatch):
+    raw_latent = torch.full((1, _LATENT_CHANNELS, 1, 4, 5), 2.0)
+    seen_shapes = []
+
+    def base_encode(self, x):
+        seen_shapes.append(tuple(x.shape))
+        return raw_latent.to(device=x.device, dtype=x.dtype)
+
+    monkeypatch.setattr(seedvr_vae_mod.VideoAutoencoderKL, "encode", base_encode)
+
+    vae = seedvr_vae_mod.VideoAutoencoderKLWrapper.__new__(seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    nn.Module.__init__(vae)
+    vae._dummy = nn.Parameter(torch.zeros((), dtype=torch.float32))
+
+    latent = vae.encode(torch.zeros(1, 3, 32, 40))
+
+    assert type(latent) is torch.Tensor
+    assert tuple(latent.shape) == (1, _LATENT_CHANNELS, 4, 5)
+    assert seen_shapes == [(1, 3, 1, 32, 40)]
+
+
+def test_seedvr2_wrapper_private_encode_helper_keeps_raw_latent(monkeypatch):
+    raw_latent = torch.full((1, _LATENT_CHANNELS, 1, 4, 5), 3.0)
+
+    def base_encode(self, x):
+        return raw_latent.to(device=x.device, dtype=x.dtype)
+
+    monkeypatch.setattr(seedvr_vae_mod.VideoAutoencoderKL, "encode", base_encode)
+
+    vae = seedvr_vae_mod.VideoAutoencoderKLWrapper.__new__(seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    nn.Module.__init__(vae)
+    vae._dummy = nn.Parameter(torch.zeros((), dtype=torch.float32))
+
+    latent, raw = vae._encode_with_raw_latent(torch.zeros(1, 3, 32, 40))
+
+    assert tuple(latent.shape) == (1, _LATENT_CHANNELS, 4, 5)
+    assert tuple(raw.shape) == (1, _LATENT_CHANNELS, 1, 4, 5)
+    assert torch.equal(raw, raw_latent)
+
+
+def _make_vae(wrapper):
+    vae = sd_mod.VAE.__new__(sd_mod.VAE)
+    vae.first_stage_model = wrapper
+    vae.device = torch.device("cpu")
+    vae.output_device = torch.device("cpu")
+    vae.vae_dtype = torch.float32
+    vae.latent_channels = _LATENT_CHANNELS
+    vae.latent_dim = 3
+    vae.downscale_ratio = (lambda a: max(0, (a + 3) // 4), 8, 8)
+    vae.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+    vae.output_channels = 3
+    vae.disable_offload = True
+    vae.extra_1d_channel = None
+    vae.crop_input = False
+    vae.not_video = False
+    vae.handles_tiling = isinstance(wrapper, seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    vae.format_encoded = wrapper.comfy_format_encoded
+    vae.patcher = _Patcher()
+    vae.process_input = lambda image: image
+    vae.process_output = lambda image: image.add(1.0).div(2.0).clamp(0.0, 1.0)
+    vae.vae_output_dtype = lambda: torch.float32
+    vae.memory_used_encode = lambda shape, dtype: 1
+    vae.memory_used_decode = lambda shape, dtype: 1
+    vae.throw_exception_if_invalid = lambda: None
+    vae.vae_encode_crop_pixels = lambda pixels: pixels
+    vae.spacial_compression_decode = lambda: 8
+    vae.temporal_compression_decode = lambda: 4
+    return vae
+
+
+def test_missing_context_falls_back_to_positive_buffer():
+    pos_buffer = torch.full((58, 5120), 7.0)
+    standin = _make_standin(pos_buffer)
+    txt, txt_shape = standin._resolve_text_conditioning(None)
+    assert txt.shape == (58, 5120)
+    assert (txt == 7.0).all(), (
+        "fallback path must use the positive_conditioning buffer "
+        "verbatim, not a zero tensor"
+    )
+    assert txt_shape.shape == (1, 1)
+    assert txt_shape[0, 0].item() == 58
+
+
+def test_seedvr2_7b_keeps_final_block_text_path(monkeypatch):
+    assert _capture_last_layer_flags(monkeypatch, vid_dim=3072, txt_in_dim=3072) == [
+        False,
+        False,
+        False,
+        False,
+    ]
+
+
+def test_seedvr2_7b_rope3d_matches_wrapper_oracle():
+    rope = seedvr_model.get_na_rope("rope3d", dim=64)
+    generator = torch.Generator(device="cpu").manual_seed(0)
+    q = torch.randn(4, 2, 128, generator=generator)
+    k = torch.randn(4, 2, 128, generator=generator)
+    shape = torch.tensor([[1, 2, 2]], dtype=torch.long)
+    freqs = rope.get_axial_freqs(1, 2, 2).reshape(4, -1)
+
+    expected_q = seedvr_model._apply_seedvr2_rotary_emb(
+        freqs,
+        q.permute(1, 0, 2).float(),
+    ).to(q.dtype).permute(1, 0, 2)
+    expected_k = seedvr_model._apply_seedvr2_rotary_emb(
+        freqs,
+        k.permute(1, 0, 2).float(),
+    ).to(k.dtype).permute(1, 0, 2)
+
+    actual_q, actual_k = rope(q.clone(), k.clone(), shape, seedvr_model.Cache(disable=True))
+
+    torch.testing.assert_close(actual_q, expected_q, rtol=0, atol=0)
+    torch.testing.assert_close(actual_k, expected_k, rtol=0, atol=0)
+
+
+def test_seedvr2_forward_requires_conditioning_latents():
+    model = NaDiT.__new__(NaDiT)
+    x = torch.zeros(1, _LATENT_CHANNELS, 1, 4, 5)
+
+    with pytest.raises(ValueError, match="requires conditioning latents"):
+        NaDiT.forward(model, x, timestep=torch.tensor([1.0]), context=None)
+
+
+def test_seedvr2_latent_format_uses_native_video_latent_shape():
+    latent_format = comfy.latent_formats.SeedVR2()
+    latent_image = torch.zeros(1, 1, 4, 5)
+
+    fixed = comfy.sample.fix_empty_latent_channels(_Model(latent_format), latent_image)
+
+    assert latent_format.latent_channels == _LATENT_CHANNELS
+    assert latent_format.latent_dimensions == 3
+    assert fixed.shape == (1, _LATENT_CHANNELS, 1, 4, 5)
+
+
+def test_seedvr2_model_requires_native_5d_latent():
+    latent = torch.zeros(1, _LATENT_CHANNELS, 2, 4, 5)
+    assert NaDiT._check_seedvr2_video_latent(latent, _LATENT_CHANNELS, "latent") is latent
+
+    with pytest.raises(ValueError, match="5-D native latent"):
+        NaDiT._check_seedvr2_video_latent(torch.zeros(1, _LATENT_CHANNELS * 2, 4, 5), _LATENT_CHANNELS, "latent")
+
+
+def test_seedvr2_encode_and_encode_tiled_preserve_native_latent_contract(monkeypatch):
+    monkeypatch.setattr(sd_mod.model_management, "load_models_gpu", lambda *a, **k: None)
+
+    encoded = torch.full((1, _LATENT_CHANNELS, 2, 4, 5), 2.0)
+    vae = _make_vae(_EncodeWrapper(encoded))
+    pixels = torch.zeros(1, 5, 32, 40, 3)
+
+    node_output = nodes_mod.VAEEncode().encode(vae, pixels)[0]
+    node_latent = node_output["samples"]
+    assert set(node_output) == {"samples"}
+    assert tuple(node_latent.shape) == (1, _LATENT_CHANNELS, 2, 4, 5)
+    assert node_latent.dtype == torch.float32
+    assert node_latent.stride()[-1] == 1
+    assert torch.equal(node_latent, torch.full_like(node_latent, 2.0 * seedvr_vae_mod.BYTEDANCE_VAE_SCALING_FACTOR))
+
+    tiled = torch.full((1, _LATENT_CHANNELS, 2, 4, 5), 3.0)
+    monkeypatch.setattr(seedvr_vae_mod, "tiled_vae", MagicMock(return_value=tiled))
+    tiled_output = nodes_mod.VAEEncodeTiled().encode(
+        vae,
+        pixels,
+        tile_size=512,
+        overlap=64,
+        temporal_size=16,
+        temporal_overlap=4,
+    )[0]
+    tiled_latent = tiled_output["samples"]
+    assert set(tiled_output) == {"samples"}
+    assert tuple(tiled_latent.shape) == (1, _LATENT_CHANNELS, 2, 4, 5)
+    assert tiled_latent.dtype == torch.float32
+    assert torch.equal(tiled_latent, torch.full_like(tiled_latent, 3.0 * seedvr_vae_mod.BYTEDANCE_VAE_SCALING_FACTOR))
+
+
+def test_vaedecode_tiled_spatial_applies_temporal_discarded(monkeypatch):
+    monkeypatch.setattr(sd_mod.model_management, "load_models_gpu", lambda *a, **k: None)
+    vae = _make_vae(_DecodeWrapper())
+
+    nodes_mod.VAEDecodeTiled().decode(
+        vae,
+        {"samples": torch.zeros(1, _LATENT_CHANNELS, 2, 4, 5)},
+        tile_size=512,
+        overlap=64,
+        temporal_size=16,
+        temporal_overlap=4,
+    )
+
+    # Spatial inputs flow through; temporal inputs are discarded as public tiling
+    # knobs, but SeedVR2's internal MemoryState causal slicing is left intact.
+    assert vae.first_stage_model.calls == [
+        {
+            "shape": (1, _LATENT_CHANNELS, 2, 4, 5),
+            "seedvr2_tiling": {
+                "enable_tiling": True,
+                "tile_size": (512, 512),
+                "tile_overlap": (64, 64),
+                "temporal_size": None,
+                "temporal_overlap": None,
+            },
+        }
+    ]

tests-unit/comfy_test/test_seedvr2_vae_decode.py

@@ -0,0 +1,94 @@
+from unittest.mock import patch
+
+import pytest
+import torch
+import torch.nn as nn
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+import comfy.ldm.seedvr.vae as vae_mod  # noqa: E402
+from comfy_extras import nodes_seedvr  # noqa: E402
+
+
+_LATENT_CHANNELS = vae_mod.SEEDVR2_LATENT_CHANNELS
+
+
+def _make_wrapper() -> vae_mod.VideoAutoencoderKLWrapper:
+    wrapper = vae_mod.VideoAutoencoderKLWrapper.__new__(
+        vae_mod.VideoAutoencoderKLWrapper
+    )
+    nn.Module.__init__(wrapper)
+    return wrapper
+
+
+def _fingerprint_decode_(self, z, return_dict=True):
+    b = int(z.shape[0])
+    t = int(z.shape[2])
+    h = int(z.shape[3])
+    w = int(z.shape[4])
+    out = torch.empty(b, 3, t, h * 8, w * 8)
+    for batch_idx in range(b):
+        out[batch_idx].fill_(float(batch_idx + 1))
+    return out
+
+
+def _decode_with_patches(wrapper, z):
+    with patch.object(vae_mod.VideoAutoencoderKL, "decode_", _fingerprint_decode_):
+        return wrapper.decode(z)
+
+
+def test_decode_b2_t3_multi_frame_batch_unchanged():
+    wrapper = _make_wrapper()
+
+    out = _decode_with_patches(wrapper, torch.zeros(2, _LATENT_CHANNELS * 3, 2, 2))
+
+    assert tuple(out.shape) == (2, 3, 3, 16, 16)
+
+
+class _Wrapper(vae_mod.VideoAutoencoderKLWrapper):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self.calls = []
+
+    def parameters(self):
+        return iter([torch.nn.Parameter(torch.zeros(()))])
+
+def _decode_stub(self, latent):
+    self.calls.append(tuple(latent.shape))
+    return torch.zeros(latent.shape[0], 3, latent.shape[2], latent.shape[3] * 8, latent.shape[4] * 8)
+
+
+def test_seedvr2_wrapper_decode_accepts_5d_channel_first_latents_without_preprocessor_state():
+    wrapper = _Wrapper()
+
+    with patch.object(vae_mod.VideoAutoencoderKL, "decode_", _decode_stub):
+        out = wrapper.decode(torch.zeros(1, _LATENT_CHANNELS, 2, 4, 5))
+
+    assert tuple(out.shape) == (1, 3, 2, 32, 40)
+    assert wrapper.calls == [(1, _LATENT_CHANNELS, 2, 4, 5)]
+
+
+def test_seedvr2_wrapper_decode_rejects_wrong_rank_latents():
+    wrapper = _Wrapper()
+
+    with pytest.raises(RuntimeError, match=r"latent input must be 4-D collapsed .* or 5-D"):
+        wrapper.decode(torch.zeros(1, _LATENT_CHANNELS, 4))
+
+
+def _t_padded(t_in: int) -> int:
+    if t_in == 1:
+        return 1
+    if t_in <= 4:
+        return 5
+    if (t_in - 1) % 4 == 0:
+        return t_in
+    return t_in + (4 - ((t_in - 1) % 4))
+
+
+@pytest.mark.parametrize("t_in", [1, 5, 9])
+def test_t_padded_matches_cut_videos(t_in):
+    dummy = torch.zeros(1, t_in, 1, 1, 1)
+    assert nodes_seedvr.cut_videos(dummy).shape[1] == _t_padded(t_in)

tests-unit/comfy_test/test_seedvr2_vae_tiled.py

@@ -0,0 +1,407 @@
+from contextlib import ExitStack
+from unittest.mock import MagicMock, patch
+
+import pytest
+import torch
+import torch.nn as nn
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+import comfy.ldm.seedvr.vae as vae_mod  # noqa: E402
+import comfy.ldm.seedvr.vae as seedvr_vae_mod  # noqa: E402
+import comfy.sd as sd_mod  # noqa: E402
+from comfy.ldm.seedvr.vae import MemoryState, tiled_vae  # noqa: E402
+
+
+_LATENT_CHANNELS = seedvr_vae_mod.SEEDVR2_LATENT_CHANNELS
+
+
+def test_runtime_decode_zero_temporal_size_preserves_model_slicing():
+    class StubVAEModel(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.slicing_latent_min_size = 2
+            self.spatial_downsample_factor = 8
+            self.temporal_downsample_factor = 4
+            self.device = torch.device("cpu")
+            self.use_slicing = True
+            self._dummy = torch.nn.Parameter(torch.zeros(1, dtype=torch.float32))
+            self.decode_min_sizes = []
+            self.memory_states = []
+
+        def decode_(self, t_chunk):
+            self.decode_min_sizes.append(self.slicing_latent_min_size)
+            return vae_mod.VideoAutoencoderKL.slicing_decode(self, t_chunk)
+
+        def _decode(self, z, memory_state=MemoryState.DISABLED, memory_cache=None):
+            self.memory_states.append(memory_state)
+            b, c, d, h, w = z.shape
+            return torch.zeros((b, 3, d, h * 8, w * 8), dtype=z.dtype)
+
+    vae = StubVAEModel()
+    z = torch.zeros((1, _LATENT_CHANNELS, 5, 8, 8), dtype=torch.float32)
+
+    tiled_vae(
+        z,
+        vae,
+        tile_size=(64, 64),
+        tile_overlap=(0, 0),
+        temporal_size=0,
+        temporal_overlap=0,
+        encode=False,
+    )
+
+    assert vae.decode_min_sizes == [2]
+    assert vae.memory_states == [MemoryState.INITIALIZING, MemoryState.ACTIVE]
+    assert vae.slicing_latent_min_size == 2
+
+
+def test_zero_temporal_size_preserves_min_size_when_encode_raises():
+    class RaisingVAEModel(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.slicing_sample_min_size = 4
+            self.spatial_downsample_factor = 8
+            self.temporal_downsample_factor = 4
+            self.device = torch.device("cpu")
+            self._dummy = torch.nn.Parameter(torch.zeros(1, dtype=torch.float32))
+
+        def encode(self, t_chunk):
+            raise RuntimeError("simulated encode failure")
+
+    vae = RaisingVAEModel()
+    x = torch.zeros((1, 3, 12, 64, 64), dtype=torch.float32)
+
+    with pytest.raises(RuntimeError, match="simulated encode failure"):
+        tiled_vae(
+            x,
+            vae,
+            tile_size=(64, 64),
+            tile_overlap=(0, 0),
+            temporal_size=0,
+            temporal_overlap=0,
+            encode=True,
+        )
+
+    assert vae.slicing_sample_min_size == 4
+
+
+def test_tiled_vae_encode_uses_tensor_return_without_indexing():
+    class TensorEncodeVAEModel(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.slicing_sample_min_size = 4
+            self.spatial_downsample_factor = 8
+            self.temporal_downsample_factor = 4
+            self.device = torch.device("cpu")
+            self._dummy = torch.nn.Parameter(torch.zeros(1, dtype=torch.float32))
+            self.calls = []
+
+        def encode(self, t_chunk):
+            self.calls.append(tuple(t_chunk.shape))
+            b, _, _, h, w = t_chunk.shape
+            return torch.ones((b, _LATENT_CHANNELS, 1, h // 8, w // 8), dtype=t_chunk.dtype)
+
+    vae = TensorEncodeVAEModel()
+    x = torch.zeros((2, 3, 1, 64, 64), dtype=torch.float32)
+
+    out = tiled_vae(
+        x,
+        vae,
+        tile_size=(64, 64),
+        tile_overlap=(0, 0),
+        temporal_size=0,
+        temporal_overlap=0,
+        encode=True,
+    )
+
+    assert vae.calls == [(2, 3, 1, 64, 64)]
+    assert tuple(out.shape) == (2, _LATENT_CHANNELS, 1, 8, 8)
+
+
+def test_tiled_vae_preserves_compute_dtype_with_different_parameter_dtype():
+    class DummyVAE(nn.Module):
+        spatial_downsample_factor = 8
+        temporal_downsample_factor = 4
+        slicing_sample_min_size = 8
+
+        def __init__(self):
+            super().__init__()
+            self.device = torch.device("cpu")
+            self._dummy = nn.Parameter(torch.zeros(1, dtype=torch.float16))
+            self.input_dtype = None
+
+        def encode(self, t_chunk):
+            self.input_dtype = t_chunk.dtype
+            b, _, _, h, w = t_chunk.shape
+            return torch.ones((b, _LATENT_CHANNELS, 1, h // 8, w // 8), dtype=t_chunk.dtype)
+
+    vae = DummyVAE()
+    x = torch.zeros((1, 3, 1, 64, 64), dtype=torch.float32)
+
+    tiled_vae(x, vae, tile_size=(64, 64), tile_overlap=(16, 16), encode=True)
+
+    assert vae.input_dtype == torch.float32
+
+
+def test_tiled_vae_preserves_input_dtype_on_single_tile():
+    class FloatOutputVAEModel(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.slicing_sample_min_size = 4
+            self.spatial_downsample_factor = 8
+            self.temporal_downsample_factor = 4
+            self.device = torch.device("cpu")
+            self._dummy = torch.nn.Parameter(torch.zeros(1, dtype=torch.float32))
+
+        def encode(self, t_chunk):
+            b, _, _, h, w = t_chunk.shape
+            return torch.ones((b, _LATENT_CHANNELS, 1, h // 8, w // 8), dtype=torch.float32)
+
+    out = tiled_vae(
+        torch.zeros((1, 3, 1, 64, 64), dtype=torch.float16),
+        FloatOutputVAEModel(),
+        tile_size=(64, 64),
+        tile_overlap=(0, 0),
+        temporal_size=0,
+        temporal_overlap=0,
+        encode=True,
+    )
+
+    assert out.dtype == torch.float16
+
+
+class _SlicingDecodeVAE(nn.Module):
+    def __init__(self, slicing_latent_min_size):
+        super().__init__()
+        self.slicing_latent_min_size = slicing_latent_min_size
+        self.spatial_downsample_factor = 8
+        self.temporal_downsample_factor = 4
+        self.device = torch.device("cpu")
+        self.use_slicing = True
+        self._dummy = nn.Parameter(torch.zeros(1, dtype=torch.float32))
+        self.decode_min_sizes = []
+        self.memory_states = []
+
+    def decode_(self, z):
+        self.decode_min_sizes.append(self.slicing_latent_min_size)
+        return vae_mod.VideoAutoencoderKL.slicing_decode(self, z)
+
+    def _decode(self, z, memory_state=MemoryState.DISABLED, memory_cache=None):
+        self.memory_states.append(memory_state)
+        x = z[:, :1].repeat(
+            1,
+            3,
+            1,
+            self.spatial_downsample_factor,
+            self.spatial_downsample_factor,
+        )
+        return x
+
+
+def test_decode_tiled_vae_maps_temporal_args_to_latent_slicing_min_size():
+    vae = _SlicingDecodeVAE(slicing_latent_min_size=2)
+    z = torch.arange(
+        _LATENT_CHANNELS * 5 * 8 * 8,
+        dtype=torch.float32,
+    ).reshape(1, _LATENT_CHANNELS, 5, 8, 8)
+
+    tiled_vae(
+        z,
+        vae,
+        tile_size=(64, 64),
+        tile_overlap=(0, 0),
+        temporal_size=12,
+        temporal_overlap=4,
+        encode=False,
+    )
+
+    assert vae.decode_min_sizes == [2]
+    assert vae.memory_states == [MemoryState.INITIALIZING, MemoryState.ACTIVE]
+    assert vae.slicing_latent_min_size == 2
+
+    wrapper = vae_mod.VideoAutoencoderKLWrapper.__new__(
+        vae_mod.VideoAutoencoderKLWrapper
+    )
+    nn.Module.__init__(wrapper)
+    seedvr2_tiling = {
+        "enable_tiling": True,
+        "tile_size": (64, 64),
+        "tile_overlap": (0, 0),
+        "temporal_size": 8,
+        "temporal_overlap": 7,
+    }
+
+    captured = {}
+
+    def _fake_tiled_vae(latent, model, **kwargs):
+        captured.update(kwargs)
+        return torch.zeros(1, 3, 1, 16, 16)
+
+    with patch.object(vae_mod, "tiled_vae", side_effect=_fake_tiled_vae):
+        wrapper.decode(torch.zeros(1, _LATENT_CHANNELS, 2, 2), seedvr2_tiling=seedvr2_tiling)
+
+    assert captured["temporal_overlap"] == 7
+
+
+def _force_oom(*a, **k):
+    raise torch.cuda.OutOfMemoryError("forced OOM for dispatcher test")
+
+
+def _make_vae(first_stage_model, latent_channels, latent_dim):
+    vae = sd_mod.VAE.__new__(sd_mod.VAE)
+    vae.first_stage_model = first_stage_model
+    vae.patcher = MagicMock()
+    vae.patcher.get_free_memory = MagicMock(return_value=8 * 1024 * 1024 * 1024)
+    vae.device = vae.output_device = torch.device("cpu")
+    vae.vae_dtype = torch.float32
+    vae.disable_offload = True
+    vae.extra_1d_channel = None
+    vae.upscale_ratio = vae.downscale_ratio = 8
+    vae.upscale_index_formula = vae.downscale_index_formula = None
+    vae.output_channels = 3
+    vae.latent_channels = latent_channels
+    vae.latent_dim = latent_dim
+    vae.vae_output_dtype = lambda: torch.float32
+    vae.spacial_compression_decode = lambda: 8
+    vae.handles_tiling = isinstance(first_stage_model, seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    vae.format_encoded = None
+    vae.process_input = lambda x: x
+    vae.process_output = lambda x: x
+    vae.throw_exception_if_invalid = lambda: None
+    vae.memory_used_decode = lambda *a, **k: 1
+    return vae
+
+
+def _dispatch(vae, samples, seedvr2_call, generic_call, patch_wrapper_decode):
+    mm = sd_mod.model_management
+    with ExitStack() as stack:
+        stack.enter_context(patch.object(mm, "raise_non_oom", lambda e: None))
+        stack.enter_context(patch.object(mm, "load_models_gpu", lambda *a, **k: None))
+        stack.enter_context(patch.object(mm, "soft_empty_cache", lambda: None))
+        stack.enter_context(patch.object(sd_mod.VAE, "_decode_tiled_owned", seedvr2_call))
+        stack.enter_context(patch.object(sd_mod.VAE, "decode_tiled_", generic_call))
+        if patch_wrapper_decode:
+            stack.enter_context(patch.object(
+                seedvr_vae_mod.VideoAutoencoderKLWrapper, "decode",
+                side_effect=_force_oom))
+        vae.decode(samples)
+
+
+def test_4d_seedvr2_latent_routes_to_owned_decode_tiled():
+    wrapper = seedvr_vae_mod.VideoAutoencoderKLWrapper.__new__(
+        seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    vae = _make_vae(wrapper, latent_channels=_LATENT_CHANNELS, latent_dim=3)
+    seedvr2_call = MagicMock(return_value=torch.zeros(1, 3, 9, 64, 64))
+    generic_call = MagicMock(return_value=torch.zeros(1, 3, 64, 64))
+    _dispatch(vae, torch.zeros(1, _LATENT_CHANNELS * 3, 8, 8), seedvr2_call, generic_call, True)
+    assert seedvr2_call.call_count == 1
+    assert generic_call.call_count == 0
+
+
+def test_4d_non_seedvr2_latent_still_routes_to_generic_decode_tiled():
+    first_stage = MagicMock()
+    first_stage.decode = MagicMock(side_effect=_force_oom)
+    vae = _make_vae(first_stage, latent_channels=4, latent_dim=2)
+    seedvr2_call = MagicMock(return_value=torch.zeros(1, 3, 9, 64, 64))
+    generic_call = MagicMock(return_value=torch.zeros(1, 3, 64, 64))
+    _dispatch(vae, torch.zeros(1, 4, 8, 8), seedvr2_call, generic_call, False)
+    assert generic_call.call_count == 1
+    assert seedvr2_call.call_count == 0
+
+
+def _populate_common_vae_attrs_fallback(vae):
+    vae.patcher = MagicMock()
+    vae.patcher.get_free_memory = MagicMock(return_value=8 * 1024 * 1024 * 1024)
+    vae.device = torch.device("cpu")
+    vae.output_device = torch.device("cpu")
+    vae.vae_dtype = torch.float32
+    vae.disable_offload = True
+    vae.extra_1d_channel = None
+    vae.upscale_ratio = 8
+    vae.upscale_index_formula = None
+    vae.output_channels = 3
+    vae.latent_channels = _LATENT_CHANNELS
+    vae.latent_dim = 3
+    vae.downscale_ratio = 8
+    vae.downscale_index_formula = None
+    vae.not_video = False
+    vae.crop_input = False
+    vae.pad_channel_value = None
+    vae.handles_tiling = isinstance(vae.first_stage_model, seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    vae.format_encoded = None
+
+    vae.vae_output_dtype = lambda: torch.float32
+    vae.spacial_compression_encode = lambda: 8
+    vae.process_input = lambda x: x
+    vae.process_output = lambda x: x
+    vae.throw_exception_if_invalid = lambda: None
+    vae.memory_used_encode = lambda *a, **k: 1
+
+
+def _make_seedvr2_vae_fallback():
+    vae = sd_mod.VAE.__new__(sd_mod.VAE)
+    wrapper = seedvr_vae_mod.VideoAutoencoderKLWrapper.__new__(
+        seedvr_vae_mod.VideoAutoencoderKLWrapper
+    )
+    vae.first_stage_model = wrapper
+    _populate_common_vae_attrs_fallback(vae)
+    return vae
+
+
+def _make_non_seedvr2_vae_fallback():
+    vae = sd_mod.VAE.__new__(sd_mod.VAE)
+    vae.first_stage_model = MagicMock()
+    _populate_common_vae_attrs_fallback(vae)
+    return vae
+
+
+def _force_regular_encode_oom(*args, **kwargs):
+    raise torch.cuda.OutOfMemoryError("forced OOM for dispatcher test")
+
+
+def test_seedvr2_3d_routes_to_owned_encode_tiled_on_oom():
+    vae = _make_seedvr2_vae_fallback()
+    pixel_samples = torch.zeros((1, 8, 64, 64, 3))
+
+    seedvr2_call = MagicMock(return_value=torch.zeros(1, _LATENT_CHANNELS, 2, 8, 8))
+    generic_call = MagicMock(return_value=torch.zeros(1, _LATENT_CHANNELS, 2, 8, 8))
+
+    with patch.object(sd_mod.model_management, "raise_non_oom",
+                      lambda e: None), \
+         patch.object(sd_mod.model_management, "load_models_gpu",
+                      lambda *a, **k: None), \
+         patch.object(sd_mod.model_management, "soft_empty_cache",
+                      lambda: None), \
+         patch.object(seedvr_vae_mod.VideoAutoencoderKLWrapper, "encode",
+                      side_effect=_force_regular_encode_oom), \
+         patch.object(sd_mod.VAE, "_encode_tiled_owned", seedvr2_call), \
+         patch.object(sd_mod.VAE, "encode_tiled_3d", generic_call):
+        vae.encode(pixel_samples)
+
+    assert seedvr2_call.call_count == 1, (
+        f"Expected _encode_tiled_owned to be called once for a SeedVR2 3D "
+        f"input under OOM fallback; got {seedvr2_call.call_count} calls."
+    )
+    assert generic_call.call_count == 0, (
+        f"encode_tiled_3d must NOT be called for a SeedVR2 input; got "
+        f"{generic_call.call_count} calls."
+    )
+
+
+def test_non_seedvr2_encode_tiled_3d_default_overlap_is_concrete():
+    vae = _make_non_seedvr2_vae_fallback()
+    vae.downscale_ratio = (lambda a: max(1, a // 4), 8, 8)
+    vae.upscale_ratio = (lambda a: a * 4, 8, 8)
+    generic_call = MagicMock(return_value=torch.zeros(1, _LATENT_CHANNELS, 2, 8, 8))
+    pixel_samples = torch.zeros((1, 8, 64, 64, 3))
+
+    with patch.object(sd_mod.model_management, "load_models_gpu",
+                      lambda *a, **k: None), \
+         patch.object(sd_mod.VAE, "encode_tiled_3d", generic_call):
+        vae.encode_tiled(pixel_samples)
+
+    assert generic_call.call_args.kwargs["overlap"] == (1, 64, 64)

tests/execution/test_execution.py

@@ -818,6 +818,30 @@ class TestExecution:
         except urllib.error.HTTPError:
             pass  # Expected behavior
 
+    def test_cached_outputs_in_job_without_client_id(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=32, width=32, batch_size=1)
+        output = g.node("SaveImage", images=image.out(0))
+
+        # Prime the cache with a normal run.
+        client.run(g)
+
+        # Resubmit anonymously (no client_id) so output nodes are cache hits with no websocket client.
+        data = json.dumps({"prompt": g.finalize()}).encode('utf-8')
+        req = urllib.request.Request(f"http://{client.server_address}/prompt", data=data)
+        prompt_id = json.loads(urllib.request.urlopen(req).read())['prompt_id']
+
+        for _ in range(100):
+            job = client.get_job(prompt_id)
+            if job is not None and job['status'] not in ('pending', 'in_progress'):
+                break
+            time.sleep(0.1)
+        else:
+            raise AssertionError("Prompt did not complete in time")
+
+        assert job['status'] == 'completed'
+        assert output.id in job['outputs'], "Cached outputs must appear in job outputs without a client_id"
+
     def _create_history_item(self, client, builder):
         g = GraphBuilder(prefix="offset_test")
         input_node = g.node(