Comfy-Org/ComfyUI · error · ValueError

SeedVR2 expected an even text-conditioning batch, got shape

Error message

SeedVR2 expected an even text-conditioning batch, got shape {tuple(context.shape)}

What it means

SeedVR2 packs positive and negative text conditioning into a single flattened stream by chunking the context batch in half (neg/pos). When cond_or_uncond indicates a mixed batch, the leading batch dimension must be even; an odd batch raises this ValueError. Single-branch batches (all cond or all uncond) are exempt via _seedvr2_is_single_conditioning_branch.

Source

Thrown at comfy/ldm/seedvr/model.py:1246

            self.vid_out_ada = ada(
                dim=vid_dim,
                emb_dim=emb_dim,
                layers=["out"],
                modes=["in"],
                device=device, dtype=dtype
            )

    def _resolve_text_conditioning(self, context, cond_or_uncond=None):
        if context is None or context.numel() == 0:
            context = self.positive_conditioning
            return flatten([context])
        if NaDiT._seedvr2_is_single_conditioning_branch(cond_or_uncond):
            if context.shape[0] == 1:
                context = context.squeeze(0)
                return flatten([context])
            return flatten(context.unbind(0))
        if context.shape[0] % 2 != 0:
            raise ValueError(f"SeedVR2 expected an even text-conditioning batch, got shape {tuple(context.shape)}")
        neg_cond, pos_cond = context.chunk(2, dim=0)
        if pos_cond.shape[0] == 1:
            pos_cond, neg_cond = pos_cond.squeeze(0), neg_cond.squeeze(0)
            return flatten([pos_cond, neg_cond])
        return flatten((*pos_cond.unbind(0), *neg_cond.unbind(0)))

    @staticmethod
    def _seedvr2_is_single_conditioning_branch(cond_or_uncond):
        if cond_or_uncond is None or len(cond_or_uncond) == 0:
            return False
        first = cond_or_uncond[0]
        return all(entry == first for entry in cond_or_uncond)

    @staticmethod
    def _check_seedvr2_video_latent(x, channels, name):
        if x.ndim != 5:
            raise ValueError(f"SeedVR2 expected {name} to be 5-D native latent, got shape {tuple(x.shape)}.")
        if x.shape[1] != channels:

View on GitHub (pinned to 1c6d8d45b3)

Solutions

Ensure cond and uncond counts are equal so context.shape[0] is even in mixed batches.
When running a single branch only, pass the correct cond_or_uncond (all zeros or all ones) so the single-branch fast path applies.
Batch symmetrically: for each positive embedding include its negative counterpart before calling forward.
In custom samplers, mirror ComfyUI's calc_cond_batch ordering and cond_or_uncond flags.

Example fix

# before
context = torch.cat([pos_a, pos_b, neg_a], 0)  # odd batch, mixed
# after
context = torch.cat([pos_a, pos_b, neg_a, neg_b], 0)  # even batch

Defensive patterns

Strategy: validation

Validate before calling

def validate_seedvr2_context(context, cond_or_uncond):
    single = cond_or_uncond is not None and len(set(cond_or_uncond)) == 1
    if not single and context.shape[0] % 2 != 0:
        raise ValueError(f"mixed cond/uncond batch must be even, got {context.shape[0]}")
    return context

Type guard

def is_single_branch(cond_or_uncond) -> bool:
    return cond_or_uncond is not None and len(cond_or_uncond) > 0 and len(set(cond_or_uncond)) == 1

Prevention

Always batch cond and uncond embeddings in equal counts.
Pass accurate cond_or_uncond flags in transformer_options for single-branch runs.
Mirror ComfyUI's calc_cond_batch ordering in custom samplers.

When it happens

Trigger: Calling NaDiT forward with a context tensor whose shape[0] is odd while transformer_options['cond_or_uncond'] contains mixed entries (e.g. [0,1]); custom samplers that concatenate cond/uncond embeddings asymmetrically; a batch of 3 where 2 are cond and 1 is uncond.

Common situations: Custom sampling loops that build conditioning batches by hand; a node that duplicates one conditioning but not its negative counterpart; partial batching where the last batch has an odd count but cond_or_uncond was not marked single-branch.

Related errors

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/b77250b76034f841. Report an issue: GitHub.