Merge 512e7767e8 into 9c34f5f36a

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Alexander Brown <DrJKL0424@gmail.com>

app/user_manager.py

@@ -28,8 +28,8 @@ def get_file_info(path: str, relative_to: str) -> FileInfo:
     return {
         "path": os.path.relpath(path, relative_to).replace(os.sep, '/'),
         "size": os.path.getsize(path),
-        "modified": os.path.getmtime(path),
-        "created": os.path.getctime(path)
+        "modified": int(os.path.getmtime(path) * 1000),
+        "created": int(os.path.getctime(path) * 1000),
     }
 
 

comfy/latent_formats.py

@@ -793,9 +793,27 @@ class ZImagePixelSpace(ChromaRadiance):
     pass
 
 class CogVideoX(LatentFormat):
+    """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b).
+
+    scale_factor matches the vae/config.json scaling_factor for the 2b variant.
+    The 5b-class checkpoints (CogVideoX-5b, CogVideoX-1.5-5B, CogVideoX-Fun-V1.5-*)
+    use a different value; see CogVideoX1_5 below.
+    """
     latent_channels = 16
     latent_dimensions = 3
     temporal_downscale_ratio = 4
 
     def __init__(self):
         self.scale_factor = 1.15258426
+
+
+class CogVideoX1_5(CogVideoX):
+    """Latent format for 5b-class CogVideoX checkpoints.
+
+    Covers THUDM/CogVideoX-5b, THUDM/CogVideoX-1.5-5B, and the CogVideoX-Fun
+    V1.5-5b family (including VOID inpainting). All of these have
+    scaling_factor=0.7 in their vae/config.json. Auto-selected in
+    supported_models.CogVideoX_T2V based on transformer hidden dim.
+    """
+    def __init__(self):
+        self.scale_factor = 0.7

comfy/ldm/twinflow/model.py

@@ -0,0 +1,763 @@
+"""
+TwinFlow-Z-Image custom model architecture for ComfyUI.
+Based on the Lumina-Image 2.0 / Z-Image architecture.
+Supports the unique dual timestep embedding architecture of TwinFlow.
+"""
+
+from __future__ import annotations
+
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import comfy.ldm.common_dit
+
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
+from comfy.ldm.modules.attention import optimized_attention_masked
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
+import comfy.patcher_extension
+
+
+def clamp_fp16(x):
+    if x.dtype == torch.float16:
+        return torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+    return x
+
+
+def modulate(x, scale):
+    return x * (1 + scale.unsqueeze(1))
+
+
+class JointAttention(nn.Module):
+    """Multi-head attention module with combined QKV weights."""
+
+    def __init__(
+        self,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: Optional[int],
+        qk_norm: bool,
+        out_bias: bool = False,
+        operation_settings={},
+    ):
+        super().__init__()
+        self.n_kv_heads = n_heads if n_kv_heads is None else n_kv_heads
+        self.n_local_heads = n_heads
+        self.n_local_kv_heads = self.n_kv_heads
+        self.n_rep = self.n_local_heads // self.n_local_kv_heads
+        self.head_dim = dim // n_heads
+
+        self.qkv = operation_settings.get("operations").Linear(
+            dim,
+            (n_heads + self.n_kv_heads + self.n_kv_heads) * self.head_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.out = operation_settings.get("operations").Linear(
+            n_heads * self.head_dim,
+            dim,
+            bias=out_bias,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        if qk_norm:
+            self.q_norm = operation_settings.get("operations").RMSNorm(
+                self.head_dim,
+                elementwise_affine=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            )
+            self.k_norm = operation_settings.get("operations").RMSNorm(
+                self.head_dim,
+                elementwise_affine=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            )
+        else:
+            self.q_norm = self.k_norm = nn.Identity()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        transformer_options={},
+    ) -> torch.Tensor:
+        bsz, seqlen, _ = x.shape
+
+        xq, xk, xv = torch.split(
+            self.qkv(x),
+            [
+                self.n_local_heads * self.head_dim,
+                self.n_local_kv_heads * self.head_dim,
+                self.n_local_kv_heads * self.head_dim,
+            ],
+            dim=-1,
+        )
+        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
+        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+
+        xq = self.q_norm(xq)
+        xk = self.k_norm(xk)
+
+        xq, xk = apply_rope(xq, xk, freqs_cis)
+
+        n_rep = self.n_local_heads // self.n_local_kv_heads
+        if n_rep > 1:
+            xk = xk.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
+            xv = xv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
+
+        output = optimized_attention_masked(
+            xq.movedim(1, 2),
+            xk.movedim(1, 2),
+            xv.movedim(1, 2),
+            self.n_local_heads,
+            x_mask,
+            skip_reshape=True,
+            transformer_options=transformer_options,
+        )
+
+        return self.out(output)
+
+
+class FeedForward(nn.Module):
+    """Feed-forward module with SiLU gating."""
+
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int,
+        ffn_dim_multiplier: Optional[float],
+        operation_settings={},
+    ):
+        super().__init__()
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = operation_settings.get("operations").Linear(
+            dim,
+            hidden_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.w2 = operation_settings.get("operations").Linear(
+            hidden_dim,
+            dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.w3 = operation_settings.get("operations").Linear(
+            dim,
+            hidden_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+    def _forward_silu_gating(self, x1, x3):
+        return clamp_fp16(F.silu(x1) * x3)
+
+    def forward(self, x):
+        return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
+
+
+class TwinFlowTransformerBlock(nn.Module):
+    """Transformer block with adaLN modulation for TwinFlow."""
+
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=False,
+        z_image_modulation=False,
+        attn_out_bias=False,
+        operation_settings={},
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.head_dim = dim // n_heads
+        self.attention = JointAttention(
+            dim,
+            n_heads,
+            n_kv_heads,
+            qk_norm,
+            out_bias=attn_out_bias,
+            operation_settings=operation_settings,
+        )
+        self.feed_forward = FeedForward(
+            dim=dim,
+            hidden_dim=dim,
+            multiple_of=multiple_of,
+            ffn_dim_multiplier=ffn_dim_multiplier,
+            operation_settings=operation_settings,
+        )
+        self.layer_id = layer_id
+        self.attention_norm1 = operation_settings.get("operations").RMSNorm(
+            dim,
+            eps=norm_eps,
+            elementwise_affine=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.ffn_norm1 = operation_settings.get("operations").RMSNorm(
+            dim,
+            eps=norm_eps,
+            elementwise_affine=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.attention_norm2 = operation_settings.get("operations").RMSNorm(
+            dim,
+            eps=norm_eps,
+            elementwise_affine=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.ffn_norm2 = operation_settings.get("operations").RMSNorm(
+            dim,
+            eps=norm_eps,
+            elementwise_affine=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        if z_image_modulation:
+            self.adaLN_modulation = nn.Sequential(
+                operation_settings.get("operations").Linear(
+                    min(dim, 256),
+                    4 * dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        adaln_input: Optional[torch.Tensor] = None,
+        transformer_options={},
+    ):
+        if adaln_input is not None:
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
+
+            x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
+                clamp_fp16(
+                    self.attention(
+                        modulate(self.attention_norm1(x), scale_msa),
+                        x_mask,
+                        freqs_cis,
+                        transformer_options=transformer_options,
+                    )
+                )
+            )
+            x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
+                clamp_fp16(
+                    self.feed_forward(
+                        modulate(self.ffn_norm1(x), scale_mlp),
+                    )
+                )
+            )
+        else:
+            x = x + self.attention_norm2(
+                clamp_fp16(
+                    self.attention(
+                        self.attention_norm1(x),
+                        x_mask,
+                        freqs_cis,
+                        transformer_options=transformer_options,
+                    )
+                )
+            )
+            x = x + self.ffn_norm2(self.feed_forward(self.ffn_norm1(x)))
+        return x
+
+
+class FinalLayer(nn.Module):
+    """Final layer with LayerNorm and output projection."""
+
+    def __init__(
+        self,
+        hidden_size: int,
+        patch_size: int,
+        out_channels: int,
+        z_image_modulation=False,
+        operation_settings={},
+    ):
+        super().__init__()
+        self.norm_final = operation_settings.get("operations").LayerNorm(
+            hidden_size,
+            elementwise_affine=False,
+            eps=1e-6,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.linear = operation_settings.get("operations").Linear(
+            hidden_size,
+            patch_size * patch_size * out_channels,
+            bias=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        min_mod = 256 if z_image_modulation else 1024
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operation_settings.get("operations").Linear(
+                min(hidden_size, min_mod),
+                hidden_size,
+                bias=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+        )
+
+    def forward(self, x: torch.Tensor, c: torch.Tensor):
+        scale = self.adaLN_modulation(c)
+        x = modulate(self.norm_final(x), scale)
+        x = self.linear(x)
+        return x
+
+
+class TwinFlowZImageTransformer(nn.Module):
+    """
+    TwinFlow-Z-Image transformer model.
+
+    This custom architecture handles dual timestep embeddings
+    (t_embedder and t_embedder_2), the primary TwinFlow distinction.
+    """
+
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        dim: int = 3840,
+        n_layers: int = 30,
+        n_refiner_layers: int = 2,
+        n_heads: int = 30,
+        n_kv_heads: Optional[int] = None,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: float = 2.6666666666666665,
+        norm_eps: float = 1e-5,
+        qk_norm: bool = True,
+        cap_feat_dim: int = 2560,
+        axes_dims: List[int] = (32, 48, 48),
+        axes_lens: List[int] = (1, 1536, 512, 512),
+        rope_theta: float = 256.0,
+        z_image_modulation: bool = True,
+        time_scale: float = 1000.0,
+        pad_tokens_multiple=None,
+        clip_text_dim=None,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {
+            "operations": operations,
+            "device": device,
+            "dtype": dtype,
+        }
+        self.time_embed_dim = 256 if z_image_modulation else min(dim, 1024)
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.patch_size = patch_size
+        self.time_scale = time_scale
+        self.pad_tokens_multiple = pad_tokens_multiple
+
+        self.x_embedder = operation_settings.get("operations").Linear(
+            in_features=patch_size * patch_size * in_channels,
+            out_features=dim,
+            bias=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        self.t_embedder = TimestepEmbedder(
+            min(dim, 1024),
+            output_size=self.time_embed_dim if z_image_modulation else None,
+            **operation_settings,
+        )
+        self.t_embedder_2 = TimestepEmbedder(
+            min(dim, 1024),
+            output_size=self.time_embed_dim if z_image_modulation else None,
+            **operation_settings,
+        )
+
+        self.noise_refiner = nn.ModuleList(
+            [
+                TwinFlowTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    modulation=True,
+                    z_image_modulation=z_image_modulation,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+
+        self.context_refiner = nn.ModuleList(
+            [
+                TwinFlowTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    modulation=False,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+
+        self.cap_embedder = nn.Sequential(
+            operation_settings.get("operations").RMSNorm(
+                cap_feat_dim,
+                eps=norm_eps,
+                elementwise_affine=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+            operation_settings.get("operations").Linear(
+                cap_feat_dim,
+                dim,
+                bias=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+        )
+
+        self.clip_text_pooled_proj = None
+        if clip_text_dim is not None:
+            self.clip_text_dim = clip_text_dim
+            self.clip_text_pooled_proj = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(
+                    clip_text_dim,
+                    eps=norm_eps,
+                    elementwise_affine=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+                operation_settings.get("operations").Linear(
+                    clip_text_dim,
+                    clip_text_dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+            self.clip_text_concat_proj = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(
+                    clip_text_dim + self.time_embed_dim,
+                    eps=norm_eps,
+                    elementwise_affine=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+                operation_settings.get("operations").Linear(
+                    clip_text_dim + self.time_embed_dim,
+                    self.time_embed_dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+
+        self.layers = nn.ModuleList(
+            [
+                TwinFlowTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    z_image_modulation=z_image_modulation,
+                    attn_out_bias=False,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_layers)
+            ]
+        )
+
+        self.final_layer = FinalLayer(
+            dim,
+            patch_size,
+            self.out_channels,
+            z_image_modulation=z_image_modulation,
+            operation_settings=operation_settings,
+        )
+
+        if self.pad_tokens_multiple is not None:
+            self.x_pad_token = nn.Parameter(torch.zeros((1, dim), device=device, dtype=dtype))
+            self.cap_pad_token = nn.Parameter(torch.zeros((1, dim), device=device, dtype=dtype))
+
+        assert (dim // n_heads) == sum(axes_dims)
+        self.axes_dims = axes_dims
+        self.axes_lens = axes_lens
+        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=rope_theta, axes_dim=axes_dims)
+        self.dim = dim
+        self.n_heads = n_heads
+
+    def _compute_twinflow_adaln(self, t: torch.Tensor, x_dtype: torch.dtype, transformer_options={}):
+        """
+        Compute TwinFlow adaLN input.
+
+        If `target_timestep` is provided in transformer options, apply the
+        TwinFlow delta-time conditioning:
+          t_emb + t_embedder_2((target - t) * time_scale) * abs(target - t)
+        otherwise fallback to the baseline additive embedding.
+        """
+        t_emb = self.t_embedder(t * self.time_scale, dtype=x_dtype)
+        target_timestep = transformer_options.get("target_timestep", None)
+        if target_timestep is None:
+            t_emb_2 = self.t_embedder_2(t * self.time_scale, dtype=x_dtype)
+            return t_emb + t_emb_2
+
+        target_t = torch.as_tensor(target_timestep, device=t.device, dtype=t.dtype)
+        if target_t.ndim == 0:
+            target_t = target_t.expand_as(t)
+
+        # If values look scaled (roughly sigma/timestep in [0..1000]), normalize.
+        t_abs_max = float(t.detach().abs().max().item()) if t.numel() else 0.0
+        tt_abs_max = float(target_t.detach().abs().max().item()) if target_t.numel() else 0.0
+        scaled_domain = (max(t_abs_max, tt_abs_max) > 2.0) and (self.time_scale > 2.0)
+        if scaled_domain:
+            t_norm = t / self.time_scale
+            tt_norm = target_t / self.time_scale
+        else:
+            t_norm = t
+            tt_norm = target_t
+
+        delta_abs = (t_norm - tt_norm).abs().unsqueeze(1).to(t_emb.dtype)
+        diff_in = (tt_norm - t_norm) * self.time_scale
+        t_emb_2 = self.t_embedder_2(diff_in, dtype=x_dtype)
+        return t_emb + t_emb_2 * delta_abs
+
+    def unpatchify(
+        self,
+        x: torch.Tensor,
+        img_size: List[Tuple[int, int]],
+        cap_size: List[int],
+        return_tensor=False,
+    ) -> List[torch.Tensor]:
+        pH = pW = self.patch_size
+        imgs = []
+        for i in range(x.size(0)):
+            H, W = img_size[i]
+            begin = cap_size[i]
+            end = begin + (H // pH) * (W // pW)
+            imgs.append(
+                x[i][begin:end]
+                .view(H // pH, W // pW, pH, pW, self.out_channels)
+                .permute(4, 0, 2, 1, 3)
+                .flatten(3, 4)
+                .flatten(1, 2)
+            )
+
+        if return_tensor:
+            imgs = torch.stack(imgs, dim=0)
+        return imgs
+
+    def patchify_and_embed(
+        self,
+        x: List[torch.Tensor] | torch.Tensor,
+        cap_feats: torch.Tensor,
+        cap_mask: torch.Tensor,
+        t: torch.Tensor,
+        num_tokens,
+        transformer_options={},
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
+        bsz = len(x)
+        pH = pW = self.patch_size
+        device = x[0].device
+
+        cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device)
+        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0
+
+        B, C, H, W = x.shape
+        x = self.x_embedder(
+            x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)
+        )
+
+        rope_options = transformer_options.get("rope_options", {})
+        h_scale = rope_options.get("scale_y", 1.0)
+        w_scale = rope_options.get("scale_x", 1.0)
+        h_start = rope_options.get("shift_y", 0.0)
+        w_start = rope_options.get("shift_x", 0.0)
+
+        H_tokens, W_tokens = H // pH, W // pW
+        x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device)
+        x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1
+        x_pos_ids[:, :, 1] = (
+            torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start
+        ).view(-1, 1).repeat(1, W_tokens).flatten()
+        x_pos_ids[:, :, 2] = (
+            torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start
+        ).view(1, -1).repeat(H_tokens, 1).flatten()
+
+        x_pad_extra = 0
+        if self.pad_tokens_multiple is not None:
+            x_pad_extra = (-x.shape[1]) % self.pad_tokens_multiple
+            x = torch.cat(
+                (
+                    x,
+                    self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], x_pad_extra, 1),
+                ),
+                dim=1,
+            )
+            x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, x_pad_extra))
+
+        cap_pad_extra = 0
+        if self.pad_tokens_multiple is not None:
+            cap_pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
+            cap_feats = torch.cat(
+                (
+                    cap_feats,
+                    self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True)
+                    .unsqueeze(0)
+                    .repeat(cap_feats.shape[0], cap_pad_extra, 1),
+                ),
+                dim=1,
+            )
+            cap_pos_ids = torch.nn.functional.pad(cap_pos_ids, (0, 0, 0, cap_pad_extra), value=0)
+            if cap_mask is not None and cap_pad_extra > 0:
+                cap_mask = torch.nn.functional.pad(cap_mask, (0, cap_pad_extra), value=0)
+
+        freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
+
+        for layer in self.context_refiner:
+            cap_feats = layer(
+                cap_feats,
+                cap_mask,
+                freqs_cis[:, : cap_pos_ids.shape[1]],
+                transformer_options=transformer_options,
+            )
+
+        padded_img_mask = None
+        for _, layer in enumerate(self.noise_refiner):
+            x = layer(
+                x,
+                padded_img_mask,
+                freqs_cis[:, cap_pos_ids.shape[1] :],
+                t,
+                transformer_options=transformer_options,
+            )
+
+        padded_full_embed = torch.cat((cap_feats, x), dim=1)
+        if cap_mask is not None:
+            cap_mask_bool = cap_mask if cap_mask.dtype == torch.bool else cap_mask > 0
+            img_mask = torch.ones((bsz, x.shape[1]), device=cap_mask.device, dtype=torch.bool)
+            if x_pad_extra > 0:
+                img_mask[:, -x_pad_extra:] = False
+            mask = torch.cat((cap_mask_bool, img_mask), dim=1)
+        else:
+            mask = None
+        img_sizes = [(H, W)] * bsz
+        l_effective_cap_len = [cap_feats.shape[1]] * bsz
+
+        return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
+
+    def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(
+                comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL,
+                kwargs.get("transformer_options", {}),
+            ),
+        ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
+
+    def _forward(
+        self,
+        x,
+        timesteps,
+        context,
+        num_tokens,
+        attention_mask=None,
+        transformer_options=None,
+        **kwargs,
+    ):
+        if transformer_options is None:
+            transformer_options = {}
+
+        t = 1.0 - timesteps
+
+        adaln_input = self._compute_twinflow_adaln(t, x.dtype, transformer_options=transformer_options)
+
+        cap_feats = context
+        cap_mask = attention_mask
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+
+        cap_feats = self.cap_embedder(cap_feats)
+
+        if self.clip_text_pooled_proj is not None:
+            pooled = kwargs.get("clip_text_pooled", None)
+            if pooled is not None:
+                pooled = self.clip_text_pooled_proj(pooled)
+            else:
+                pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype)
+            adaln_input = torch.cat((adaln_input, pooled), dim=-1)
+            adaln_input = self.clip_text_concat_proj(adaln_input)
+
+        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(
+            x,
+            cap_feats,
+            cap_mask,
+            adaln_input,
+            num_tokens,
+            transformer_options=transformer_options,
+        )
+        freqs_cis = freqs_cis.to(img.device)
+
+        transformer_options["total_blocks"] = len(self.layers)
+        transformer_options["block_type"] = "double"
+
+        for i, layer in enumerate(self.layers):
+            transformer_options["block_index"] = i
+            img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
+
+        img = self.final_layer(img, adaln_input)
+        img = self.unpatchify(
+            img,
+            img_size,
+            cap_size,
+            return_tensor=isinstance(x, torch.Tensor),
+        )[:, :, :h, :w]
+
+        return -img

comfy/lora.py

@@ -329,6 +329,15 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to
                 key_map[key_lora] = to
 
+    # TwinFlow-Z-Image LoRAs can target t_embedder_2.* keys.
+    # Alias them back to t_embedder.* targets for compatibility.
+    if isinstance(model, comfy.model_base.TwinFlow_Z_Image):
+        for key in list(key_map.keys()):
+            if "t_embedder." in key and "t_embedder_2." not in key:
+                key_2 = key.replace("t_embedder.", "t_embedder_2.", 1)
+                if key_2 not in key_map:
+                    key_map[key_2] = key_map[key]
+                
     if isinstance(model, comfy.model_base.Kandinsky5):
         for k in sdk:
             if k.startswith("diffusion_model.") and k.endswith(".weight"):

comfy/lora_convert.py

@@ -32,7 +32,15 @@ def convert_uso_lora(sd):
         sd_out[k_to] = tensor
     return sd_out
 
-
+def twinflow_z_image_lora_to_diffusers(state_dict):
+    """Convert TwinFlow LoRA state dict for diffusers compatibility."""
+    for key in list(state_dict.keys()):
+        if "t_embedder_2" not in key and key.startswith("t_embedder."):
+            new_key = key.replace("t_embedder.", "t_embedder_2.", 1)
+            if new_key not in state_dict:
+                state_dict[new_key] = state_dict.pop(key)
+    return state_dict
+    
 def convert_lora(sd):
     if "img_in.lora_A.weight" in sd and "single_blocks.0.norm.key_norm.scale" in sd:
         return convert_lora_bfl_control(sd)
@@ -40,4 +48,6 @@ def convert_lora(sd):
         return convert_lora_wan_fun(sd)
     if "single_blocks.37.processor.qkv_lora.up.weight" in sd and "double_blocks.18.processor.qkv_lora2.up.weight" in sd:
         return convert_uso_lora(sd)
+    if any(k.startswith("t_embedder.") for k in sd.keys()):
+        return twinflow_z_image_lora_to_diffusers(sd)
     return sd

comfy/model_base.py

@@ -40,6 +40,7 @@ import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
 import comfy.ldm.cosmos.predict2
 import comfy.ldm.lumina.model
+import comfy.ldm.twinflow.model
 import comfy.ldm.wan.model
 import comfy.ldm.wan.model_animate
 import comfy.ldm.wan.ar_model
@@ -1293,6 +1294,11 @@ class ZImagePixelSpace(Lumina2):
         BaseModel.__init__(self, model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiTPixelSpace)
         self.memory_usage_factor_conds = ("ref_latents",)
 
+class TwinFlow_Z_Image(Lumina2):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        BaseModel.__init__(self, model_config, model_type, device=device, unet_model=comfy.ldm.twinflow.model.TwinFlowZImageTransformer)
+        self.memory_usage_factor_conds = ("ref_latents",)
+        
 class WAN21(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)

comfy/model_detection.py

@@ -44,6 +44,48 @@ def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
 def detect_unet_config(state_dict, key_prefix, metadata=None):
     state_dict_keys = list(state_dict.keys())
 
+    # TwinFlow-Z-Image: detect dual timestep embedder checkpoints first.
+    if any(k.startswith('{}t_embedder_2.'.format(key_prefix)) for k in state_dict_keys):
+        dit_config = {
+            "image_model": "twinflow_z_image",
+            "architecture": "TwinFlow_Z_Image",
+            "patch_size": 2,
+            "in_channels": 16,
+            "qk_norm": True,
+            "ffn_dim_multiplier": (8.0 / 3.0),
+            "z_image_modulation": True,
+            "time_scale": 1000.0,
+            "n_refiner_layers": 2,
+        }
+
+        cap_embedder_key = '{}cap_embedder.1.weight'.format(key_prefix)
+        if cap_embedder_key in state_dict:
+            w = state_dict[cap_embedder_key]
+            dit_config["dim"] = w.shape[0]
+            dit_config["cap_feat_dim"] = w.shape[1]
+
+        dit_config["n_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.')
+
+        # Match Z-Image style defaults (TwinFlow checkpoints are 3840-dim variants).
+        dit_config["n_heads"] = 30
+        dit_config["n_kv_heads"] = 30
+        dit_config["axes_dims"] = [32, 48, 48]
+        dit_config["axes_lens"] = [1536, 512, 512]
+        dit_config["rope_theta"] = 256.0
+
+        try:
+            dit_config["allow_fp16"] = torch.std(
+                state_dict['{}layers.{}.ffn_norm1.weight'.format(key_prefix, dit_config["n_layers"] - 2)],
+                unbiased=False
+            ).item() < 0.42
+        except Exception:
+            pass
+
+        if '{}cap_pad_token'.format(key_prefix) in state_dict_keys or '{}x_pad_token'.format(key_prefix) in state_dict_keys:
+            dit_config["pad_tokens_multiple"] = 32
+
+        return dit_config
+        
     if '{}joint_blocks.0.context_block.attn.qkv.weight'.format(key_prefix) in state_dict_keys: #mmdit model
         unet_config = {}
         unet_config["in_channels"] = state_dict['{}x_embedder.proj.weight'.format(key_prefix)].shape[1]

comfy/sd.py

@@ -66,6 +66,7 @@ import comfy.text_encoders.longcat_image
 import comfy.text_encoders.qwen35
 import comfy.text_encoders.ernie
 import comfy.text_encoders.gemma4
+import comfy.text_encoders.cogvideo
 
 import comfy.model_patcher
 import comfy.lora
@@ -78,6 +79,20 @@ import comfy.latent_formats
 
 import comfy.ldm.flux.redux
 
+def is_twinflow_z_image_model(state_dict):
+    """Check if model state dict is TwinFlow-Z-Image."""
+    return any(k.startswith("t_embedder_2.") for k in state_dict)
+
+
+def get_twinflow_z_image_config(state_dict):
+    """Extract TwinFlow-Z-Image configuration from state dict."""
+    if not is_twinflow_z_image_model(state_dict):
+        return {}
+    return {
+        "image_model": "twinflow_z_image",
+        "architecture": "TwinFlow_Z_Image",
+    }
+    
 def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
     key_map = {}
     if model is not None:
@@ -1224,6 +1239,7 @@ class CLIPType(Enum):
     NEWBIE = 24
     FLUX2 = 25
     LONGCAT_IMAGE = 26
+    COGVIDEOX = 27
 
 
 
@@ -1428,6 +1444,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
+            elif clip_type == CLIPType.COGVIDEOX:
+                clip_target.clip = comfy.text_encoders.cogvideo.cogvideo_te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.cogvideo.CogVideoXTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer

comfy/supported_models.py

@@ -1134,6 +1134,15 @@ class ZImagePixelSpace(ZImage):
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.ZImagePixelSpace(self, device=device)
 
+class TwinFlow_Z_Image(ZImage):
+    unet_config = {
+        "image_model": "twinflow_z_image",
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.TwinFlow_Z_Image(self, device=device)
+        return out
+        
 class WAN21_T2V(supported_models_base.BASE):
     unet_config = {
         "image_model": "wan2.1",
@@ -1872,6 +1881,14 @@ class CogVideoX_T2V(supported_models_base.BASE):
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
+    def __init__(self, unet_config):
+        # 2b-class (dim=1920, heads=30) uses scale_factor=1.15258426.
+        # 5b-class (dim=3072, heads=48) — incl. CogVideoX-5b, 1.5-5B, and
+        # Fun-V1.5 inpainting — uses scale_factor=0.7 per vae/config.json.
+        if unet_config.get("num_attention_heads", 0) >= 48:
+            self.latent_format = latent_formats.CogVideoX1_5
+        super().__init__(unet_config)
+
     def get_model(self, state_dict, prefix="", device=None):
         # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE
         if self.unet_config.get("patch_size_t") is not None:
@@ -1898,6 +1915,20 @@ class CogVideoX_I2V(CogVideoX_T2V):
         out = model_base.CogVideoX(self, image_to_video=True, device=device)
         return out
 
+class CogVideoX_Inpaint(CogVideoX_T2V):
+    unet_config = {
+        "image_model": "cogvideox",
+        "in_channels": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, image_to_video=True, device=device)
+        return out
+
 
 models = [
     LotusD,
@@ -1946,6 +1977,7 @@ models = [
     CosmosI2VPredict2,
     ZImagePixelSpace,
     ZImage,
+    TwinFlow_Z_Image,
     Lumina2,
     WAN22_T2V,
     WAN21_CausalAR_T2V,
@@ -1978,6 +2010,7 @@ models = [
     ErnieImage,
     SAM3,
     SAM31,
+    CogVideoX_Inpaint,
     CogVideoX_I2V,
     CogVideoX_T2V,
     SVD_img2vid,

comfy/text_encoders/cogvideo.py

@@ -1,6 +1,48 @@
 import comfy.text_encoders.sd3_clip
+from comfy import sd1_clip
 
 
 class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer):
+    """Inner T5 tokenizer for CogVideoX.
+
+    CogVideoX was trained with T5 embeddings padded to 226 tokens (not 77 like SD3).
+    Used both directly by supported_models.CogVideoX_T2V.clip_target (paired with
+    the raw T5XXLModel) and by the CogVideoXTokenizer outer wrapper below.
+    """
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226)
+
+
+class CogVideoXTokenizer(sd1_clip.SD1Tokenizer):
+    """Outer tokenizer wrapper for CLIPLoader (type="cogvideox")."""
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data,
+                         clip_name="t5xxl", tokenizer=CogVideoXT5Tokenizer)
+
+
+class CogVideoXT5XXL(sd1_clip.SD1ClipModel):
+    """Outer T5XXL model wrapper for CLIPLoader (type="cogvideox").
+
+    Wraps the raw T5XXL model in the SD1ClipModel interface so that CLIP.__init__
+    (which reads self.dtypes) works correctly. The inner model is the standard
+    sd3_clip.T5XXLModel (no attention_mask change needed for CogVideoX).
+    """
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl",
+                         clip_model=comfy.text_encoders.sd3_clip.T5XXLModel,
+                         model_options=model_options)
+
+
+def cogvideo_te(dtype_t5=None, t5_quantization_metadata=None):
+    """Factory that returns a CogVideoXT5XXL class configured with the detected
+    T5 dtype and optional quantization metadata, for use in load_text_encoder_state_dicts.
+    """
+    class CogVideoXTEModel_(CogVideoXT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
+            if dtype_t5 is not None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return CogVideoXTEModel_

comfy/utils.py

@@ -818,6 +818,17 @@ def z_image_to_diffusers(mmdit_config, output_prefix=""):
 
     return key_map
 
+def twinflow_z_image_key_mapping(state_dict, key):
+    """
+    TwinFlow-Z-Image key mapping.
+    Maps t_embedder_2 keys to t_embedder for weight loading.
+    """
+    if key.startswith("t_embedder_2."):
+        new_key = key.replace("t_embedder_2.", "t_embedder.", 1)
+        if new_key not in state_dict:
+            state_dict[new_key] = state_dict.pop(key)
+    return state_dic
+
 def repeat_to_batch_size(tensor, batch_size, dim=0):
     if tensor.shape[dim] > batch_size:
         return tensor.narrow(dim, 0, batch_size)

comfy_extras/nodes_ace.py

@@ -42,7 +42,7 @@ class TextEncodeAceStepAudio15(IO.ComfyNode):
                 IO.Int.Input("bpm", default=120, min=10, max=300),
                 IO.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1),
                 IO.Combo.Input("timesignature", options=['2', '3', '4', '6']),
-                IO.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
+                IO.Combo.Input("language", options=['ar', 'az', 'bg', 'bn', 'ca', 'cs', 'da', 'de', 'el', 'en', 'es', 'fa', 'fi', 'fr', 'he', 'hi', 'hr', 'ht', 'hu', 'id', 'is', 'it', 'ja', 'ko', 'la', 'lt', 'ms', 'ne', 'nl', 'no', 'pa', 'pl', 'pt', 'ro', 'ru', 'sa', 'sk', 'sr', 'sv', 'sw', 'ta', 'te', 'th', 'tl', 'tr', 'uk', 'ur', 'vi', 'yue', 'zh', 'unknown'], default='en'),
                 IO.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
                 IO.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True),
                 IO.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True),

comfy_extras/nodes_void.py

@@ -0,0 +1,483 @@
+import logging
+
+import torch
+
+import comfy
+import comfy.model_management
+import comfy.model_patcher
+import comfy.samplers
+import comfy.utils
+import folder_paths
+import node_helpers
+import nodes
+from comfy.utils import model_trange as trange
+from comfy_api.latest import ComfyExtension, io
+from torchvision.models.optical_flow import raft_large
+from typing_extensions import override
+
+
+from comfy_extras.void_noise_warp import RaftOpticalFlow, get_noise_from_video
+
+OpticalFlow = io.Custom("OPTICAL_FLOW")
+
+TEMPORAL_COMPRESSION = 4
+PATCH_SIZE_T = 2
+
+
+def _valid_void_length(length: int) -> int:
+    """Round ``length`` down to a value that produces an even latent_t.
+
+    VOID / CogVideoX-Fun-V1.5 uses patch_size_t=2, so the VAE-encoded latent
+    must have an even temporal dimension. If latent_t is odd, the transformer
+    pad_to_patch_size circular-wraps an extra latent frame onto the end; after
+    the post-transformer crop the last real latent frame has been influenced
+    by the wrapped phantom frame, producing visible jitter and "disappearing"
+    subjects near the end of the decoded video. Rounding down fixes this.
+    """
+    latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+    if latent_t % PATCH_SIZE_T == 0:
+        return length
+    # Round latent_t down to the nearest multiple of PATCH_SIZE_T, then invert
+    # the ((length - 1) // TEMPORAL_COMPRESSION) + 1 formula. Floor at 1 frame
+    # so we never return a non-positive length.
+    target_latent_t = max(PATCH_SIZE_T, (latent_t // PATCH_SIZE_T) * PATCH_SIZE_T)
+    return (target_latent_t - 1) * TEMPORAL_COMPRESSION + 1
+
+
+class OpticalFlowLoader(io.ComfyNode):
+    """Load an optical flow model from ``models/optical_flow/``.
+
+    Only torchvision's RAFT-large format is recognized today (the model used
+    by VOIDWarpedNoise).  The checkpoint must be placed under
+    ``models/optical_flow/`` — ComfyUI never downloads optical-flow weights
+    at runtime.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="OpticalFlowLoader",
+            display_name="Load Optical Flow Model",
+            category="loaders",
+            inputs=[
+                io.Combo.Input(
+                    "model_name",
+                    options=folder_paths.get_filename_list("optical_flow"),
+                    tooltip=(
+                        "Optical flow model to load.  Files must be placed in the "
+                        "'optical_flow' folder.  Today only torchvision's "
+                        "raft_large.pth is supported."
+                    ),
+                ),
+            ],
+            outputs=[
+                OpticalFlow.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_name) -> io.NodeOutput:
+
+        model_path = folder_paths.get_full_path_or_raise("optical_flow", model_name)
+        sd = comfy.utils.load_torch_file(model_path, safe_load=True)
+
+        has_raft_keys = (
+            any(k.startswith("feature_encoder.") for k in sd)
+            and any(k.startswith("context_encoder.") for k in sd)
+            and any(k.startswith("update_block.") for k in sd)
+        )
+        if not has_raft_keys:
+            raise ValueError(
+                "Unrecognized optical flow model format: expected a torchvision "
+                "RAFT-large state dict with 'feature_encoder.', 'context_encoder.' "
+                "and 'update_block.' prefixes."
+            )
+
+        model = raft_large(weights=None, progress=False)
+        model.load_state_dict(sd)
+        model.eval().to(torch.float32)
+
+        patcher = comfy.model_patcher.ModelPatcher(
+            model,
+            load_device=comfy.model_management.get_torch_device(),
+            offload_device=comfy.model_management.unet_offload_device(),
+        )
+        return io.NodeOutput(patcher)
+
+
+class VOIDQuadmaskPreprocess(io.ComfyNode):
+    """Preprocess a quadmask video for VOID inpainting.
+
+    Quantizes mask values to four semantic levels, inverts, and normalizes:
+      0   -> primary object to remove
+      63  -> overlap of primary + affected
+      127 -> affected region (interactions)
+      255 -> background (keep)
+
+    After inversion and normalization, the output mask has values in [0, 1]
+    with four discrete levels: 1.0 (remove), ~0.75, ~0.50, 0.0 (keep).
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDQuadmaskPreprocess",
+            category="mask/video",
+            inputs=[
+                io.Mask.Input("mask"),
+                io.Int.Input("dilate_width", default=0, min=0, max=50, step=1,
+                             tooltip="Dilation radius for the primary mask region (0 = no dilation)"),
+            ],
+            outputs=[
+                io.Mask.Output(display_name="quadmask"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, mask, dilate_width=0) -> io.NodeOutput:
+        m = mask.clone()
+
+        if m.max() <= 1.0:
+            m = m * 255.0
+
+        if dilate_width > 0 and m.ndim >= 3:
+            binary = (m < 128).float()
+            kernel_size = dilate_width * 2 + 1
+            if binary.ndim == 3:
+                binary = binary.unsqueeze(1)
+            dilated = torch.nn.functional.max_pool2d(
+                binary, kernel_size=kernel_size, stride=1, padding=dilate_width
+            )
+            if dilated.ndim == 4:
+                dilated = dilated.squeeze(1)
+            m = torch.where(dilated > 0.5, torch.zeros_like(m), m)
+
+        m = torch.where(m <= 31, torch.zeros_like(m), m)
+        m = torch.where((m > 31) & (m <= 95), torch.full_like(m, 63), m)
+        m = torch.where((m > 95) & (m <= 191), torch.full_like(m, 127), m)
+        m = torch.where(m > 191, torch.full_like(m, 255), m)
+
+        m = (255.0 - m) / 255.0
+
+        return io.NodeOutput(m)
+
+
+class VOIDInpaintConditioning(io.ComfyNode):
+    """Build VOID inpainting conditioning for CogVideoX.
+
+    Encodes the processed quadmask and masked source video through the VAE,
+    producing a 32-channel concat conditioning (16ch mask + 16ch masked video)
+    that gets concatenated with the 16ch noise latent by the model.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDInpaintConditioning",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Image.Input("video", tooltip="Source video frames [T, H, W, 3]"),
+                io.Mask.Input("quadmask", tooltip="Preprocessed quadmask from VOIDQuadmaskPreprocess [T, H, W]"),
+                io.Int.Input("width", default=672, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("length", default=45, min=1, max=nodes.MAX_RESOLUTION, step=1,
+                             tooltip="Number of pixel frames to process. For CogVideoX-Fun-V1.5 "
+                                     "(patch_size_t=2), latent_t must be even — lengths that "
+                                     "produce odd latent_t are rounded down (e.g. 49 → 45)."),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, video, quadmask,
+                width, height, length, batch_size) -> io.NodeOutput:
+
+        adjusted_length = _valid_void_length(length)
+        if adjusted_length != length:
+            logging.warning(
+                "VOIDInpaintConditioning: rounding length %d down to %d so that "
+                "latent_t is even (required by CogVideoX-Fun-V1.5 patch_size_t=2). "
+                "Using odd latent_t causes the last frame to be corrupted by "
+                "circular padding.", length, adjusted_length,
+            )
+            length = adjusted_length
+
+        latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+        latent_h = height // 8
+        latent_w = width // 8
+
+        vid = video[:length]
+        vid = comfy.utils.common_upscale(
+            vid.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+
+        qm = quadmask[:length]
+        if qm.ndim == 3:
+            qm = qm.unsqueeze(-1)
+        qm = comfy.utils.common_upscale(
+            qm.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+        if qm.ndim == 4 and qm.shape[-1] == 1:
+            qm = qm.squeeze(-1)
+
+        mask_condition = qm
+        if mask_condition.ndim == 3:
+            mask_condition_3ch = mask_condition.unsqueeze(-1).expand(-1, -1, -1, 3)
+        else:
+            mask_condition_3ch = mask_condition
+
+        inverted_mask_3ch = 1.0 - mask_condition_3ch
+        masked_video = vid[:, :, :, :3] * (1.0 - mask_condition_3ch)
+
+        mask_latents = vae.encode(inverted_mask_3ch)
+        masked_video_latents = vae.encode(masked_video)
+
+        def _match_temporal(lat, target_t):
+            if lat.shape[2] > target_t:
+                return lat[:, :, :target_t]
+            elif lat.shape[2] < target_t:
+                pad = target_t - lat.shape[2]
+                return torch.cat([lat, lat[:, :, -1:].repeat(1, 1, pad, 1, 1)], dim=2)
+            return lat
+
+        mask_latents = _match_temporal(mask_latents, latent_t)
+        masked_video_latents = _match_temporal(masked_video_latents, latent_t)
+
+        inpaint_latents = torch.cat([mask_latents, masked_video_latents], dim=1)
+
+        # No explicit scaling needed here: the model's CogVideoX.concat_cond()
+        # applies process_latent_in (×latent_format.scale_factor) to each 16-ch
+        # block of the stored conditioning. For 5b-class checkpoints (incl. the
+        # VOID/CogVideoX-Fun-V1.5 inpainting model) that scale_factor is auto-
+        # selected as 0.7 in supported_models.CogVideoX_T2V, which matches the
+        # diffusers vae/config.json scaling_factor VOID was trained with.
+
+        positive = node_helpers.conditioning_set_values(
+            positive, {"concat_latent_image": inpaint_latents}
+        )
+        negative = node_helpers.conditioning_set_values(
+            negative, {"concat_latent_image": inpaint_latents}
+        )
+
+        noise_latent = torch.zeros(
+            [batch_size, 16, latent_t, latent_h, latent_w],
+            device=comfy.model_management.intermediate_device()
+        )
+
+        return io.NodeOutput(positive, negative, {"samples": noise_latent})
+
+
+class VOIDWarpedNoise(io.ComfyNode):
+    """Generate optical-flow warped noise for VOID Pass 2 refinement.
+
+    Takes the Pass 1 output video and produces temporally-correlated noise
+    by warping Gaussian noise along optical flow vectors. This noise is used
+    as the initial latent for Pass 2, resulting in better temporal consistency.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDWarpedNoise",
+            category="latent/video",
+            inputs=[
+                OpticalFlow.Input(
+                    "optical_flow",
+                    tooltip="Optical flow model from OpticalFlowLoader (RAFT-large).",
+                ),
+                io.Image.Input("video", tooltip="Pass 1 output video frames [T, H, W, 3]"),
+                io.Int.Input("width", default=672, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("length", default=45, min=1, max=nodes.MAX_RESOLUTION, step=1,
+                             tooltip="Number of pixel frames. Rounded down to make latent_t "
+                                     "even (patch_size_t=2 requirement), e.g. 49 → 45."),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Latent.Output(display_name="warped_noise"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, optical_flow, video, width, height, length, batch_size) -> io.NodeOutput:
+
+        adjusted_length = _valid_void_length(length)
+        if adjusted_length != length:
+            logging.warning(
+                "VOIDWarpedNoise: rounding length %d down to %d so that "
+                "latent_t is even (required by CogVideoX-Fun-V1.5 patch_size_t=2).",
+                length, adjusted_length,
+            )
+            length = adjusted_length
+
+        latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+        latent_h = height // 8
+        latent_w = width // 8
+
+        # RAFT + noise warp is real compute, not an "intermediate" buffer, so
+        # we want the actual torch device (CUDA/MPS).  The final latent is
+        # moved back to intermediate_device() before returning to match the
+        # rest of the ComfyUI pipeline.
+        device = comfy.model_management.get_torch_device()
+
+        comfy.model_management.load_model_gpu(optical_flow)
+        raft = RaftOpticalFlow(optical_flow.model, device=device)
+
+        vid = video[:length].to(device)
+        vid = comfy.utils.common_upscale(
+            vid.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+        vid_uint8 = (vid.clamp(0, 1) * 255).to(torch.uint8)
+
+        FRAME = 2**-1
+        FLOW = 2**3
+        LATENT_SCALE = 8
+
+        warped = get_noise_from_video(
+            vid_uint8,
+            raft,
+            noise_channels=16,
+            resize_frames=FRAME,
+            resize_flow=FLOW,
+            downscale_factor=round(FRAME * FLOW) * LATENT_SCALE,
+            device=device,
+        )
+
+        if warped.shape[0] != latent_t:
+            indices = torch.linspace(0, warped.shape[0] - 1, latent_t,
+                                     device=device).long()
+            warped = warped[indices]
+
+        if warped.shape[1] != latent_h or warped.shape[2] != latent_w:
+            # (T, H, W, C) → (T, C, H, W) → bilinear resize → back
+            warped = warped.permute(0, 3, 1, 2)
+            warped = torch.nn.functional.interpolate(
+                warped, size=(latent_h, latent_w),
+                mode="bilinear", align_corners=False,
+            )
+            warped = warped.permute(0, 2, 3, 1)
+
+        # (T, H, W, C) → (B, C, T, H, W)
+        warped_tensor = warped.permute(3, 0, 1, 2).unsqueeze(0)
+        if batch_size > 1:
+            warped_tensor = warped_tensor.repeat(batch_size, 1, 1, 1, 1)
+
+        warped_tensor = warped_tensor.to(comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": warped_tensor})
+
+
+class Noise_FromLatent:
+    """Wraps a pre-computed LATENT tensor as a NOISE source."""
+    def __init__(self, latent_dict):
+        self.seed = 0
+        self._samples = latent_dict["samples"]
+
+    def generate_noise(self, input_latent):
+        return self._samples.clone().cpu()
+
+
+class VOIDWarpedNoiseSource(io.ComfyNode):
+    """Convert a LATENT (e.g. from VOIDWarpedNoise) into a NOISE source
+    for use with SamplerCustomAdvanced."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDWarpedNoiseSource",
+            category="sampling/custom_sampling/noise",
+            inputs=[
+                io.Latent.Input("warped_noise",
+                    tooltip="Warped noise latent from VOIDWarpedNoise"),
+            ],
+            outputs=[io.Noise.Output()],
+        )
+
+    @classmethod
+    def execute(cls, warped_noise) -> io.NodeOutput:
+        return io.NodeOutput(Noise_FromLatent(warped_noise))
+
+
+class VOID_DDIM(comfy.samplers.Sampler):
+    """DDIM sampler for VOID inpainting models.
+
+    VOID was trained with the diffusers CogVideoXDDIMScheduler which operates in
+    alpha-space (input std ≈ 1). The standard KSampler applies noise_scaling that
+    multiplies by sqrt(1+sigma^2) ≈ 4500x, which is incompatible with VOID's
+    training. This sampler skips noise_scaling and implements the DDIM update rule
+    directly using sigma-to-alpha conversion.
+    """
+
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        x = noise.to(torch.float32)
+        model_options = extra_args.get("model_options", {})
+        seed = extra_args.get("seed", None)
+        s_in = x.new_ones([x.shape[0]])
+
+        for i in trange(len(sigmas) - 1, disable=disable_pbar):
+            sigma = sigmas[i]
+            sigma_next = sigmas[i + 1]
+
+            denoised = model_wrap(x, sigma * s_in, model_options=model_options, seed=seed)
+
+            if callback is not None:
+                callback(i, denoised, x, len(sigmas) - 1)
+
+            if sigma_next == 0:
+                x = denoised
+            else:
+                alpha_t = 1.0 / (1.0 + sigma ** 2)
+                alpha_prev = 1.0 / (1.0 + sigma_next ** 2)
+
+                pred_eps = (x - (alpha_t ** 0.5) * denoised) / (1.0 - alpha_t) ** 0.5
+                x = (alpha_prev ** 0.5) * denoised + (1.0 - alpha_prev) ** 0.5 * pred_eps
+
+        return x
+
+
+class VOIDSampler(io.ComfyNode):
+    """VOID DDIM sampler for use with SamplerCustom / SamplerCustomAdvanced.
+
+    Required for VOID inpainting models. Implements the same DDIM loop that VOID
+    was trained with (diffusers CogVideoXDDIMScheduler), without the noise_scaling
+    that the standard KSampler applies. Use with RandomNoise or VOIDWarpedNoiseSource.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDSampler",
+            category="sampling/custom_sampling/samplers",
+            inputs=[],
+            outputs=[io.Sampler.Output()],
+        )
+
+    @classmethod
+    def execute(cls) -> io.NodeOutput:
+        return io.NodeOutput(VOID_DDIM())
+
+    get_sampler = execute
+
+
+class VOIDExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            OpticalFlowLoader,
+            VOIDQuadmaskPreprocess,
+            VOIDInpaintConditioning,
+            VOIDWarpedNoise,
+            VOIDWarpedNoiseSource,
+            VOIDSampler,
+        ]
+
+
+async def comfy_entrypoint() -> VOIDExtension:
+    return VOIDExtension()

comfy_extras/void_noise_warp.py

@@ -0,0 +1,494 @@
+"""
+Optical-flow-warped noise for VOID Pass 2 refinement.
+
+Adapted from RyannDaGreat/CommonSource (MIT License, Ryan Burgert):
+  https://github.com/RyannDaGreat/CommonSource
+  - noise_warp.py  (NoiseWarper / warp_xyωc / regaussianize / get_noise_from_video)
+  - raft.py        (RaftOpticalFlow)
+
+Only the code paths that ``comfy_extras/nodes_void.py::VOIDWarpedNoise`` actually
+uses (torch THWC uint8 input, no background removal, no visualization, no disk
+I/O, default warp/noise params) have been inlined.  External ``rp`` utilities
+have been replaced with equivalents from torch.nn.functional / einops.  The
+RAFT optical-flow model itself is loaded offline via ``OpticalFlowLoader`` in
+``nodes_void.py`` and passed into ``get_noise_from_video`` by the caller; this
+module never downloads weights at runtime.
+"""
+
+import logging
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+import comfy.model_management
+
+
+# ---------------------------------------------------------------------------
+# Low-level torch image helpers (drop-in replacements for rp.torch_* primitives)
+# ---------------------------------------------------------------------------
+
+def _torch_resize_chw(image, size, interp, copy=True):
+    """Resize a CHW tensor.
+
+    ``size`` is either a scalar factor or a (h, w) tuple.  ``interp`` is one
+    of ``"bilinear"``, ``"nearest"``, ``"area"``.  When ``copy`` is False and
+    the requested size matches the input, returns the input tensor as is
+    (faster but callers must not mutate the result).
+    """
+    if image.ndim != 3:
+        raise ValueError(
+            f"_torch_resize_chw expects a 3D CHW tensor, got shape {tuple(image.shape)}"
+        )
+    _, in_h, in_w = image.shape
+    if isinstance(size, (int, float)) and not isinstance(size, bool):
+        new_h = max(1, int(in_h * size))
+        new_w = max(1, int(in_w * size))
+    else:
+        new_h, new_w = size
+
+    if (new_h, new_w) == (in_h, in_w):
+        return image.clone() if copy else image
+
+    kwargs = {}
+    if interp in ("bilinear", "bicubic"):
+        kwargs["align_corners"] = False
+    out = F.interpolate(image[None], size=(new_h, new_w), mode=interp, **kwargs)[0]
+    return out
+
+
+def _torch_remap_relative(image, dx, dy, interp="bilinear"):
+    """Relative remap of a CHW image via ``F.grid_sample``.
+
+    Equivalent to ``rp.torch_remap_image(image, dx, dy, relative=True, interp=interp)``
+    for ``interp`` in {"bilinear", "nearest"}.  Out-of-bounds samples are 0.
+    """
+    if image.ndim != 3:
+        raise ValueError(
+            f"_torch_remap_relative expects a 3D CHW tensor, got shape {tuple(image.shape)}"
+        )
+    if dx.shape != dy.shape:
+        raise ValueError(
+            f"_torch_remap_relative: dx and dy must match, got {tuple(dx.shape)} vs {tuple(dy.shape)}"
+        )
+    _, h, w = image.shape
+
+    x_abs = dx + torch.arange(w, device=dx.device, dtype=dx.dtype)
+    y_abs = dy + torch.arange(h, device=dy.device, dtype=dy.dtype)[:, None]
+
+    x_norm = (x_abs / (w - 1)) * 2 - 1
+    y_norm = (y_abs / (h - 1)) * 2 - 1
+
+    grid = torch.stack([x_norm, y_norm], dim=-1)[None].to(image.dtype)
+    out = F.grid_sample(
+        image[None], grid, mode=interp, align_corners=True, padding_mode="zeros"
+    )[0]
+    return out
+
+
+def _torch_scatter_add_relative(image, dx, dy):
+    """Scatter-add a CHW image using relative floor-rounded (dx, dy) offsets.
+
+    Equivalent to ``rp.torch_scatter_add_image(image, dx, dy, relative=True,
+    interp='floor')``.  Out-of-bounds targets are dropped.
+    """
+    if image.ndim != 3:
+        raise ValueError(
+            f"_torch_scatter_add_relative expects a 3D CHW tensor, got shape {tuple(image.shape)}"
+        )
+    in_c, in_h, in_w = image.shape
+    if dx.shape != (in_h, in_w) or dy.shape != (in_h, in_w):
+        raise ValueError(
+            f"_torch_scatter_add_relative: dx/dy must be ({in_h}, {in_w}), "
+            f"got dx={tuple(dx.shape)} dy={tuple(dy.shape)}"
+        )
+
+    x = dx.long() + torch.arange(in_w, device=dx.device, dtype=torch.long)
+    y = dy.long() + torch.arange(in_h, device=dy.device, dtype=torch.long)[:, None]
+
+    valid = ((y >= 0) & (y < in_h) & (x >= 0) & (x < in_w)).reshape(-1)
+    indices = (y * in_w + x).reshape(-1)[valid]
+
+    flat_image = rearrange(image, "c h w -> (h w) c")[valid]
+    out = torch.zeros((in_h * in_w, in_c), dtype=image.dtype, device=image.device)
+    out.index_add_(0, indices, flat_image)
+    return rearrange(out, "(h w) c -> c h w", h=in_h, w=in_w)
+
+
+# ---------------------------------------------------------------------------
+# Noise warping primitives (ported from noise_warp.py)
+# ---------------------------------------------------------------------------
+
+def unique_pixels(image):
+    """Find unique pixel values in a CHW tensor.
+
+    Returns ``(unique_colors [U, C], counts [U], index_matrix [H, W])`` where
+    ``index_matrix[i, j]`` is the index of the unique color at that pixel.
+    """
+    _, h, w = image.shape
+    flat = rearrange(image, "c h w -> (h w) c")
+    unique_colors, inverse_indices, counts = torch.unique(
+        flat, dim=0, return_inverse=True, return_counts=True, sorted=False,
+    )
+    index_matrix = rearrange(inverse_indices, "(h w) -> h w", h=h, w=w)
+    return unique_colors, counts, index_matrix
+
+
+def sum_indexed_values(image, index_matrix):
+    """For each unique index, sum the CHW image values at its pixels."""
+    _, h, w = image.shape
+    u = int(index_matrix.max().item()) + 1
+    flat = rearrange(image, "c h w -> (h w) c")
+    out = torch.zeros((u, flat.shape[1]), dtype=flat.dtype, device=flat.device)
+    out.index_add_(0, index_matrix.view(-1), flat)
+    return out
+
+
+def indexed_to_image(index_matrix, unique_colors):
+    """Build a CHW image from an index matrix and a (U, C) color table."""
+    h, w = index_matrix.shape
+    flat = unique_colors[index_matrix.view(-1)]
+    return rearrange(flat, "(h w) c -> c h w", h=h, w=w)
+
+
+def regaussianize(noise):
+    """Variance-preserving re-sampling of a CHW noise tensor.
+
+    Wherever the noise contains groups of identical pixel values (e.g. after
+    a nearest-neighbor warp that duplicated source pixels), adds zero-mean
+    foreign noise within each group and scales by ``1/sqrt(count)`` so the
+    output is unit-variance gaussian again.
+    """
+    _, hs, ws = noise.shape
+    _, counts, index_matrix = unique_pixels(noise[:1])
+
+    foreign_noise = torch.randn_like(noise)
+    summed = sum_indexed_values(foreign_noise, index_matrix)
+    meaned = indexed_to_image(index_matrix, summed / rearrange(counts, "u -> u 1"))
+    zeroed_foreign = foreign_noise - meaned
+
+    counts_image = indexed_to_image(index_matrix, rearrange(counts, "u -> u 1"))
+
+    output = noise / counts_image ** 0.5 + zeroed_foreign
+    return output, counts_image
+
+
+def xy_meshgrid_like_image(image):
+    """Return a (2, H, W) tensor of (x, y) pixel coordinates matching ``image``."""
+    _, h, w = image.shape
+    y, x = torch.meshgrid(
+        torch.arange(h, device=image.device, dtype=image.dtype),
+        torch.arange(w, device=image.device, dtype=image.dtype),
+        indexing="ij",
+    )
+    return torch.stack([x, y])
+
+
+def noise_to_state(noise):
+    """Pack a (C, H, W) noise tensor into a state tensor (3+C, H, W) = [dx, dy, ω, noise]."""
+    zeros = torch.zeros_like(noise[:1])
+    ones = torch.ones_like(noise[:1])
+    return torch.cat([zeros, zeros, ones, noise])
+
+
+def state_to_noise(state):
+    """Unpack the noise channels from a state tensor."""
+    return state[3:]
+
+
+def warp_state(state, flow):
+    """Warp a noise-warper state tensor along the given optical flow.
+
+    ``state`` has shape ``(3+c, h, w)`` (= dx, dy, ω, c noise channels).
+    ``flow`` has shape ``(2, h, w)`` (= dx, dy).
+    """
+    if flow.device != state.device:
+        raise ValueError(
+            f"warp_state: flow and state must be on the same device, "
+            f"got flow={flow.device} state={state.device}"
+        )
+    if state.ndim != 3:
+        raise ValueError(
+            f"warp_state: state must be 3D (3+C, H, W), got shape {tuple(state.shape)}"
+        )
+    xyoc, h, w = state.shape
+    if flow.shape != (2, h, w):
+        raise ValueError(
+            f"warp_state: flow must have shape (2, {h}, {w}), got {tuple(flow.shape)}"
+        )
+    device = state.device
+
+    x_ch, y_ch = 0, 1
+    xy = 2         # state[:xy]  = [dx, dy]
+    xyw = 3        # state[:xyw] = [dx, dy, ω]
+    w_ch = 2       # state[w_ch] = ω
+    c = xyoc - xyw
+    oc = xyoc - xy
+    if c <= 0:
+        raise ValueError(
+            f"warp_state: state has no noise channels (expected 3+C with C>0, got {xyoc} channels)"
+        )
+    if not (state[w_ch] > 0).all():
+        raise ValueError("warp_state: all weights in state[2] must be > 0")
+
+    grid = xy_meshgrid_like_image(state)
+
+    init = torch.empty_like(state)
+    init[:xy] = 0
+    init[w_ch] = 1
+    init[-c:] = 0
+
+    # --- Expansion branch: nearest-neighbor remap with negated flow ---
+    pre_expand = torch.empty_like(state)
+    pre_expand[:xy] = _torch_remap_relative(state[:xy], -flow[0], -flow[1], "nearest")
+    pre_expand[-oc:] = _torch_remap_relative(state[-oc:], -flow[0], -flow[1], "nearest")
+    pre_expand[w_ch][pre_expand[w_ch] == 0] = 1
+
+    # --- Shrink branch: scatter-add state into new positions ---
+    pre_shrink = state.clone()
+    pre_shrink[:xy] += flow
+
+    pos = (grid + pre_shrink[:xy]).round()
+    in_bounds = (pos[x_ch] >= 0) & (pos[x_ch] < w) & (pos[y_ch] >= 0) & (pos[y_ch] < h)
+    pre_shrink = torch.where(~in_bounds[None], init, pre_shrink)
+
+    scat_xy = pre_shrink[:xy].round()
+    pre_shrink[:xy] -= scat_xy
+    pre_shrink[:xy] = 0  # xy_mode='none' in upstream
+
+    def scat(tensor):
+        return _torch_scatter_add_relative(tensor, scat_xy[0], scat_xy[1])
+
+    # rp.torch_scatter_add_image on a bool tensor errors on modern torch;
+    # scatter-sum a float ones tensor and threshold to get the mask instead.
+    shrink_mask = scat(torch.ones(1, h, w, dtype=state.dtype, device=device)) > 0
+
+    # Drop expansion samples at positions that will be filled by shrink.
+    pre_expand = torch.where(shrink_mask, init, pre_expand)
+
+    # Regaussianize both branches together so duplicated-source groups are
+    # counted globally, then split back apart.
+    concat = torch.cat([pre_shrink, pre_expand], dim=2)  # along width
+    concat[-c:], counts_image = regaussianize(concat[-c:])
+    concat[w_ch] = concat[w_ch] / counts_image[0]
+    concat[w_ch] = concat[w_ch].nan_to_num()
+    pre_shrink, expand = torch.chunk(concat, chunks=2, dim=2)
+
+    shrink = torch.empty_like(pre_shrink)
+    shrink[w_ch] = scat(pre_shrink[w_ch][None])[0]
+    shrink[:xy] = scat(pre_shrink[:xy] * pre_shrink[w_ch][None]) / shrink[w_ch][None]
+    shrink[-c:] = scat(pre_shrink[-c:] * pre_shrink[w_ch][None]) / scat(
+        pre_shrink[w_ch][None] ** 2
+    ).sqrt()
+
+    output = torch.where(shrink_mask, shrink, expand)
+    output[w_ch] = output[w_ch] / output[w_ch].mean()
+    output[w_ch] += 1e-5
+    output[w_ch] **= 0.9999
+    return output
+
+
+class NoiseWarper:
+    """Maintain a warpable noise state and emit gaussian noise per frame.
+
+    Simplified from RyannDaGreat/CommonSource/noise_warp.py::NoiseWarper:
+    ``scale_factor``, ``post_noise_alpha``, ``progressive_noise_alpha``, and
+    ``warp_kwargs`` are all dropped since VOIDWarpedNoise always uses defaults.
+    """
+
+    def __init__(self, c, h, w, device, dtype=torch.float32):
+        if c <= 0 or h <= 0 or w <= 0:
+            raise ValueError(
+                f"NoiseWarper: c/h/w must all be positive, got c={c} h={h} w={w}"
+            )
+        self.c = c
+        self.h = h
+        self.w = w
+        self.device = device
+        self.dtype = dtype
+
+        noise = torch.randn(c, h, w, dtype=dtype, device=device)
+        self._state = noise_to_state(noise)
+
+    @property
+    def noise(self):
+        # With scale_factor=1 the "downsample to respect weights" step is a
+        # size-preserving no-op; the weight-variance correction math still
+        # runs to stay faithful to upstream.
+        n = state_to_noise(self._state)
+        weights = self._state[2:3]
+        return n * weights / (weights ** 2).sqrt()
+
+    def __call__(self, dx, dy):
+        if dx.shape != dy.shape:
+            raise ValueError(
+                f"NoiseWarper: dx and dy must match, got {tuple(dx.shape)} vs {tuple(dy.shape)}"
+            )
+        flow = torch.stack([dx, dy]).to(self.device, self.dtype)
+        _, oflowh, ofloww = flow.shape
+
+        flow = _torch_resize_chw(flow, (self.h, self.w), "bilinear", copy=True)
+        flowh, floww = flow.shape[-2:]
+
+        # Upstream scales flow[0] by flowh/oflowh and flow[1] by floww/ofloww
+        # (channel-order appears swapped but harmless when H and W are scaled
+        # by the same factor, which is always the case for our callers).
+        flow[0] *= flowh / oflowh
+        flow[1] *= floww / ofloww
+
+        self._state = warp_state(self._state, flow)
+        return self
+
+
+# ---------------------------------------------------------------------------
+# RAFT optical flow wrapper (ported from raft.py)
+# ---------------------------------------------------------------------------
+
+class RaftOpticalFlow:
+    """RAFT-large wrapper around a pre-loaded torchvision model.
+
+    ``model`` must be the ``torchvision.models.optical_flow.raft_large`` module
+    with its weights already populated; this class is load-agnostic so the
+    caller owns downloading/offload concerns (see ``OpticalFlowLoader`` in
+    ``nodes_void.py``).  ``__call__`` returns a ``(2, H, W)`` flow.
+    """
+
+    def __init__(self, model, device=None):
+        if device is None:
+            device = comfy.model_management.get_torch_device()
+        device = torch.device(device) if not isinstance(device, torch.device) else device
+
+        model = model.to(device)
+        model.eval()
+        self.device = device
+        self.model = model
+
+    def _preprocess(self, image_chw):
+        image = image_chw.to(self.device, torch.float32)
+        _, h, w = image.shape
+        new_h = (h // 8) * 8
+        new_w = (w // 8) * 8
+        image = _torch_resize_chw(image, (new_h, new_w), "bilinear", copy=False)
+        image = image * 2 - 1
+        return image[None]
+
+    def __call__(self, from_image, to_image):
+        """``from_image``, ``to_image``: CHW float tensors in [0, 1]."""
+        if from_image.shape != to_image.shape:
+            raise ValueError(
+                f"RaftOpticalFlow: from_image and to_image must match, "
+                f"got {tuple(from_image.shape)} vs {tuple(to_image.shape)}"
+            )
+        _, h, w = from_image.shape
+        with torch.no_grad():
+            img1 = self._preprocess(from_image)
+            img2 = self._preprocess(to_image)
+            list_of_flows = self.model(img1, img2)
+            flow = list_of_flows[-1][0]  # (2, new_h, new_w)
+            if flow.shape[-2:] != (h, w):
+                flow = _torch_resize_chw(flow, (h, w), "bilinear", copy=False)
+        return flow
+
+
+# ---------------------------------------------------------------------------
+# Narrow entry point used by VOIDWarpedNoise
+# ---------------------------------------------------------------------------
+
+def get_noise_from_video(
+    video_frames: torch.Tensor,
+    raft: RaftOpticalFlow,
+    *,
+    noise_channels: int = 16,
+    resize_frames: float = 0.5,
+    resize_flow: int = 8,
+    downscale_factor: int = 32,
+    device: Optional[torch.device] = None,
+) -> torch.Tensor:
+    """Produce optical-flow-warped gaussian noise from a video.
+
+    Args:
+        video_frames: ``(T, H, W, 3)`` uint8 torch tensor.
+        raft: Pre-loaded RAFT optical-flow wrapper (see ``RaftOpticalFlow``).
+        noise_channels: Channels in the output noise.
+        resize_frames: Pre-RAFT frame scale factor.
+        resize_flow: Post-flow up-scale factor applied to the optical flow;
+            the internal noise state is allocated at
+            ``(resize_flow * resize_frames * H, resize_flow * resize_frames * W)``.
+        downscale_factor: Area-pool factor applied to the noise before return;
+            should evenly divide the internal noise resolution.
+        device: Target device.  Defaults to ``comfy.model_management.get_torch_device()``.
+
+    Returns:
+        ``(T, H', W', noise_channels)`` float32 noise tensor on ``device``.
+    """
+    if not isinstance(resize_flow, int) or resize_flow < 1:
+        raise ValueError(
+            f"get_noise_from_video: resize_flow must be a positive int, got {resize_flow!r}"
+        )
+    if video_frames.ndim != 4 or video_frames.shape[-1] != 3:
+        raise ValueError(
+            "get_noise_from_video: video_frames must have shape (T, H, W, 3), "
+            f"got {tuple(video_frames.shape)}"
+        )
+    if video_frames.dtype != torch.uint8:
+        raise TypeError(
+            "get_noise_from_video: video_frames must be uint8 in [0, 255], "
+            f"got dtype {video_frames.dtype}"
+        )
+
+    if device is None:
+        device = comfy.model_management.get_torch_device()
+    device = torch.device(device) if not isinstance(device, torch.device) else device
+
+    if device.type == "cpu":
+        logging.warning(
+            "VOIDWarpedNoise: running get_noise_from_video on CPU; this will be "
+            "slow (minutes for ~45 frames).  Use CUDA for interactive use."
+        )
+
+    T = video_frames.shape[0]
+    frames = video_frames.to(device).permute(0, 3, 1, 2).to(torch.float32) / 255.0
+    if resize_frames != 1.0:
+        new_h = max(1, int(frames.shape[2] * resize_frames))
+        new_w = max(1, int(frames.shape[3] * resize_frames))
+        frames = F.interpolate(frames, size=(new_h, new_w), mode="area")
+
+    _, _, H, W = frames.shape
+    internal_h = resize_flow * H
+    internal_w = resize_flow * W
+    if internal_h % downscale_factor or internal_w % downscale_factor:
+        logging.warning(
+            "VOIDWarpedNoise: internal noise size %dx%d is not divisible by "
+            "downscale_factor %d; output noise may have artifacts.",
+            internal_h, internal_w, downscale_factor,
+        )
+
+    with torch.no_grad():
+        warper = NoiseWarper(
+            c=noise_channels, h=internal_h, w=internal_w, device=device,
+        )
+        down_h = warper.h // downscale_factor
+        down_w = warper.w // downscale_factor
+        output = torch.empty(
+            (T, down_h, down_w, noise_channels), dtype=torch.float32, device=device,
+        )
+
+        def downscale(noise_chw):
+            # Area-pool to 1/downscale_factor then multiply by downscale_factor
+            # to adjust std (sqrt of pool area == downscale_factor for a
+            # square pool).
+            down = _torch_resize_chw(noise_chw, 1.0 / downscale_factor, "area", copy=False)
+            return down * downscale_factor
+
+        output[0] = downscale(warper.noise).permute(1, 2, 0)
+
+        prev = frames[0]
+        for i in range(1, T):
+            curr = frames[i]
+            flow = raft(prev, curr).to(device)
+            warper(flow[0], flow[1])
+            output[i] = downscale(warper.noise).permute(1, 2, 0)
+            prev = curr
+
+    return output

folder_paths.py

@@ -54,6 +54,8 @@ folder_names_and_paths["audio_encoders"] = ([os.path.join(models_dir, "audio_enc
 
 folder_names_and_paths["frame_interpolation"] = ([os.path.join(models_dir, "frame_interpolation")], supported_pt_extensions)
 
+folder_names_and_paths["optical_flow"] = ([os.path.join(models_dir, "optical_flow")], supported_pt_extensions)
+
 output_directory = os.path.join(base_path, "output")
 temp_directory = os.path.join(base_path, "temp")
 input_directory = os.path.join(base_path, "input")

nodes.py

@@ -958,7 +958,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image", "cogvideox"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -968,7 +968,7 @@ class CLIPLoader:
 
     CATEGORY = "advanced/loaders"
 
-    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B"
+    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncogvideox: t5 xxl (226-token padding)\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B"
 
     def load_clip(self, clip_name, type="stable_diffusion", device="default"):
         clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION)
@@ -2430,6 +2430,7 @@ async def init_builtin_extra_nodes():
         "nodes_rtdetr.py",
         "nodes_frame_interpolation.py",
         "nodes_sam3.py",
+        "nodes_void.py",
     ]
 
     import_failed = []

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.42.15
+comfyui-frontend-package==1.43.17
 comfyui-workflow-templates==0.9.69
 comfyui-embedded-docs==0.4.4
 torch

tests-unit/prompt_server_test/user_manager_test.py

@@ -69,7 +69,11 @@ async def test_listuserdata_full_info(aiohttp_client, app, tmp_path):
     assert len(result) == 1
     assert result[0]["path"] == "file1.txt"
     assert "size" in result[0]
-    assert "modified" in result[0]
+    assert isinstance(result[0]["modified"], int)
+    assert isinstance(result[0]["created"], int)
+    # Verify millisecond magnitude (timestamps after year 2000 in ms are > 946684800000)
+    assert result[0]["modified"] > 946684800000
+    assert result[0]["created"] > 946684800000
 
 
 async def test_listuserdata_split_path(aiohttp_client, app, tmp_path):