Comfy-Org/ComfyUI · error · ValueError

The generated keyframes were recorded at {generated_keyframe

Error message

The generated keyframes were recorded at {generated_keyframes['tokens_per_frame']} tokens per latent frame but this latent has {tokens_per_frame}. Separate the generated keyframes before upscaling the latent.

What it means

generated_keyframes carries the tokens_per_frame it was recorded with; when the current latent's tokens_per_latent_frame differs (different spatial resolution), the frame-slot boolean built from it would mark wrong token ranges, so the model refuses. Same family as errors 174/175: keyframe bookkeeping is resolution-bound.

Source

Thrown at comfy/ldm/lightricks/model.py:1202

        return (orig_shape[3] // patch_size[1]) * (orig_shape[4] // patch_size[2])

    def keyframes_abs_pos_mask(self, pixel_coords, orig_shape, grid_mask, num_guide_tokens, generated_keyframes):
        """Per-token mask selecting the latents that encode a single standalone pixel frame.

        Returns a (batch, tokens) boolean mask over the already grid-filtered token sequence.
        """
        temporal_start = pixel_coords[:, 0]
        if temporal_start.ndim == 3:  # (batch, tokens, [start, end])
            temporal_start = temporal_start[..., 0]
        mask = temporal_start == 0
        if num_guide_tokens > 0:
            mask[:, -num_guide_tokens:] = False

        if generated_keyframes is not None:
            # The temporal patch size is always 1, so one latent frame is one row of tokens.
            tokens_per_frame = self.tokens_per_latent_frame(orig_shape)
            if generated_keyframes["tokens_per_frame"] != tokens_per_frame:
                raise ValueError(
                    f"The generated keyframes were recorded at {generated_keyframes['tokens_per_frame']} tokens "
                    f"per latent frame but this latent has {tokens_per_frame}. Separate the generated keyframes "
                    "before upscaling the latent."
                )
            first_token = generated_keyframes["first_latent_frame"] * tokens_per_frame
            num_slot_tokens = generated_keyframes["num_keyframes"] * tokens_per_frame
            slots = torch.zeros(orig_shape[2] * tokens_per_frame, dtype=torch.bool, device=mask.device)
            slots[first_token:first_token + num_slot_tokens] = True
            if grid_mask is not None:
                slots = slots[grid_mask]
            mask = mask | slots

        return mask

    def apply_keyframes_abs_pos_embedding(self, x, pixel_coords, orig_shape, grid_mask, num_guide_tokens, generated_keyframes):
        """Add the learned keyframe marker to the single-pixel-frame tokens.

        A no-op for every checkpoint built without the parameter.

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Remove generated_keyframes (separate them) before the upscaling sampling pass
2. Or regenerate the generated_keyframes dict against the upscaled latent's resolution, with first_latent_frame/num_keyframes recomputed
3. Keep a single resolution for the whole keyframe-conditioned chain

tpf = model.tokens_per_latent_frame(orig_shape)
assert generated_keyframes['tokens_per_frame'] == tpf, (generated_keyframes['tokens_per_frame'], tpf)

When it happens

Trigger: Passing generated_keyframes from a base-resolution pass into an upsampled pass where each latent frame has 4x the tokens (2x spatial upscale).

Common situations: LTX upscale pipelines that keep generated-keyframe conditioning wired across the resolution boundary.

Related errors

• keyframe_idxs holds {keyframe_idxs.shape[2]} tokens, which i
• Either spatial_upsample or temporal_upsample must be True
• guide pre_filter_counts ({total_pfc}) != keyframe grid mask
• Unrecognized interpolation method '{method}'.
• DurationHead requires at least one of video_tokens / audio_t

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/257bb2d4c7b646e0. Report an issue: GitHub.