Comfy-Org/ComfyUI · error · ValueError

SeedVR2 patch input temporal size must satisfy T % {t} == 1,

Error message

SeedVR2 patch input temporal size must satisfy T % {t} == 1, got {vid.size(2)}.

What it means

SeedVR2's PatchIn keeps the very first frame unpatched and patches groups of t consecutive frames (it duplicates frame 0 t-1 times before reshaping), so the input temporal length T must satisfy T % t == 1. A video whose frame count does not leave remainder 1 modulo the temporal patch size cannot be reshaped, hence the ValueError. This is an input-length contract on the video latent frames.

Source

Thrown at comfy/ldm/seedvr/model.py:951

        self,
        in_channels: int,
        patch_size: Union[int, Tuple[int, int, int]],
        dim: int,
        device, dtype, operations
    ):
        super().__init__()
        t, h, w = _triple(patch_size)
        self.patch_size = t, h, w
        self.proj = operations.Linear(in_channels * t * h * w, dim, device=device, dtype=dtype)

    def forward(
        self,
        vid: torch.Tensor,
    ) -> torch.Tensor:
        t, h, w = self.patch_size
        if t > 1:
            if vid.size(2) % t != 1:
                raise ValueError(
                    f"SeedVR2 patch input temporal size must satisfy T % {t} == 1, got {vid.size(2)}."
                )
            vid = torch.cat([vid[:, :, :1]] * (t - 1) + [vid], dim=2)
        b, c, Tt, Hh, Ww = vid.shape
        vid = vid.view(b, c, Tt // t, t, Hh // h, h, Ww // w, w).permute(0, 2, 4, 6, 3, 5, 7, 1).reshape(b, Tt // t, Hh // h, Ww // w, t * h * w * c)
        vid = self.proj(vid)
        return vid

class NaPatchIn(PatchIn):
    def forward(
        self,
        vid: torch.Tensor,  # l c
        vid_shape: torch.LongTensor,
        cache: Optional[Cache] = None,
    ) -> torch.Tensor:
        if cache is None:
            cache = Cache(disable=True)
        cache = cache.namespace("patch")

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Adjust the frame count so T % t == 1 (for t=2 that means an odd frame count: 1, 3, 5, ...).
2. Drop or duplicate trailing frames in preprocessing to reach a valid length, e.g. keep the first (T // t) * t + 1 frames.
3. Check the model config's patch_size triple to learn t and validate before loading the video.
4. If frames come from a VAE encode of raw video, fix the count before encoding so latent frames align.

Example fix

# before
vid = load_video(path)  # T=16, t=2 -> raises
# after
t = 2
T_valid = (vid.size(2) // t) * t + 1  # 17 -> still even split issue; instead trim to odd
vid = vid[:, :, :T_valid] if T_valid <= vid.size(2) else torch.cat([vid, vid[:, :, -1:].repeat(1, 1, T_valid - vid.size(2), 1, 1)], dim=2)

def validate_seedvr2_temporal(vid, patch_t):
    T = vid.size(2)
    if T % patch_t != 1:
        target = (T // patch_t) * patch_t + 1
        raise ValueError(f"frame count {T} invalid for patch_t={patch_t}; trim/pad to {target} frames")
    return vid

def temporal_ok(T, t) -> bool:
    return T % t == 1

When it happens

Trigger: Feeding SeedVR2 a video latent with a frame count that is a multiple of t (or arbitrary) — e.g. T=16 frames with temporal patch size 2 (16 % 2 == 0, not 1). Common with arbitrary frame counts produced by trimming, looping, or frame-interpolation preprocessing.

Common situations: Loading a video with an even frame count into a model with temporal patch size 2; trimming frames to a round number like 32; frame-count arithmetic in a preprocessing pipeline that ignores the model's T % t == 1 contract.

Related errors

• SeedVR2 expected {name} to be 5-D native latent, got shape {
• SeedVR2 conditioning shape must match latent batch/temporal/
• ar_video sampler requires 5-D video latents [B,C,T,H,W], got
• unknown merge strategy {self.merge_strategy}
• PixDiT_T2I requires context (text embeddings) of shape [B, L

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/54df064d16015d68. Report an issue: GitHub.