Comfy-Org/ComfyUI · error · ValueError

SeedVR2 conditioning shape must match latent batch/temporal/

Error message

SeedVR2 conditioning shape must match latent batch/temporal/spatial dimensions; got latent {tuple(x.shape)} and conditioning {tuple(conditions.shape)}.

What it means

After channel/ndim checks, SeedVR2 requires the conditioning latent to match the noisy latent in batch, temporal, and spatial dims (conditions.shape[0] == b and conditions.shape[2:] == (t, h, w)) because they are concatenated token-wise after flattening. Any mismatch — different resolution, frame count, or batch — aborts with both shapes printed.

Source

Thrown at comfy/ldm/seedvr/model.py:1292

    def forward(
        self,
        x,
        timestep,
        context,  # l c
        disable_cache: bool = False,
        **kwargs
    ):
        transformer_options = kwargs.get("transformer_options", {})
        patches_replace = transformer_options.get("patches_replace", {})
        blocks_replace = patches_replace.get("dit", {})
        conditions = kwargs.get("condition")
        if conditions is None:
            raise ValueError("SeedVR2 requires conditioning latents from the SeedVR2Conditioning node.")
        x = self._check_seedvr2_video_latent(x, SEEDVR2_LATENT_CHANNELS, "latent")
        conditions = self._check_seedvr2_video_latent(conditions, SEEDVR2_LATENT_CHANNELS + 1, "conditioning")
        b, _, t, h, w = x.shape
        if conditions.shape[0] != b or conditions.shape[2:] != (t, h, w):
            raise ValueError(
                f"SeedVR2 conditioning shape must match latent batch/temporal/spatial dimensions; got latent {tuple(x.shape)} and conditioning {tuple(conditions.shape)}."
            )
        x = x.movedim(1, -1)
        conditions = conditions.movedim(1, -1)
        cache = Cache(disable=disable_cache)

        txt, txt_shape = self._resolve_text_conditioning(context, transformer_options.get("cond_or_uncond"))

        vid, vid_shape = flatten(x)
        cond_latent, _ = flatten(conditions)

        vid = torch.cat([vid, cond_latent], dim=-1)

        txt = self.txt_in(txt)

        vid_shape_before_patchify = vid_shape
        vid, vid_shape = self.vid_in(vid, vid_shape, cache=cache)

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Make the LQ conditioning video exactly the target latent's T/H/W (match resolution and frame count before VAE encode).
2. Ensure batch dims agree — replicate the conditioning if you batch the latent.
3. Apply the same temporal trimming (T % t == 1) to both latent and conditioning.
4. Use the SeedVR2Conditioning node, which validates alignment, instead of manual concatenation.

Example fix

# before
cond = lq_latent  # encoded at 720x1280, latent is 724x1284
# after
lq = torch.nn.functional.interpolate(lq_video, size=(T, H_out, W_out), mode='trilinear')
cond = seedvr_vae.encode(lq)  # same T,H,W as target latent

def check_conditioning_alignment(x, cond):
    b, _, t, h, w = x.shape
    if cond.shape[0] != b or cond.shape[2:] != (t, h, w):
        raise ValueError(
            f"conditioning {tuple(cond.shape)} must match latent (b={b}, t={t}, h={h}, w={w})"
        )
    return cond

def conditioning_aligned(x, cond) -> bool:
    return cond.dim() == 5 and cond.shape[0] == x.shape[0] and cond.shape[2:] == x.shape[2:]

When it happens

Trigger: Encoding the LQ/reference video at a different resolution or frame count than the generation target; batch size of the conditioning differing from the latent batch; cropping/trimming one input but not the other.

Common situations: Upscaling workflows where the LQ video was resized to the target resolution with slightly different rounding (off-by-one dims); frame-count changes after trimming to satisfy T % t == 1 on only one side; batch-2 sampling with a batch-1 conditioning latent.

Related errors

• PixDiT_T2I requires context (text embeddings) of shape [B, L
• SeedVR2 patch input temporal size must satisfy T % {t} == 1,
• SeedVR2 expected an even text-conditioning batch, got shape
• SeedVR2 expected {name} to be 5-D native latent, got shape {
• SeedVR2 expected {name} channels to be {channels}, got shape

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/565ef5881fbc859f. Report an issue: GitHub.