fix: video frame validation to enable automatic clipping in inference pipeline

[jeanachoi]

Issues

1. The error message does not clearly mention that the input must have at least 121 (self.pixel_chunk_duration, this is a set value) frames.
2. Validation occurs too early in the pipeline, stopping videos from reaching the clipping logic that handles partial chunks. Currently, get_pixel_num_frames and get_latent_num_frames raise errors if the frame count isn't divisible by 121. This prevents inputs with more than 121 frames that aren't multiples of 121 from proceeding, even though autoclipping logic exists further down the pipeline before tensor reshaping.

Fix
Validation is already enforced where tensor reshaping happens.

cosmos-transfer1/cosmos_transfer1/diffusion/module/pretrained_vae.py

Lines 361 to 376 in bc18a6b

	def transform_encode_state_shape(self, state: torch.Tensor) -> torch.Tensor:
	"""
	Rearranges the input state tensor to the required shape for encoding video data. Mainly for chunk based encoding
	"""
	B, C, T, H, W = state.shape
	assert (
	T % self.pixel_chunk_duration == 0
	), f"Temporal dimension {T} is not divisible by chunk_length {self.pixel_chunk_duration}"
	return rearrange(state, "b c (n t) h w -> (b n) c t h w", t=self.pixel_chunk_duration)
	def transform_decode_state_shape(self, latent: torch.Tensor) -> torch.Tensor:
	B, _, T, _, _ = latent.shape
	assert (
	T % self.latent_chunk_duration == 0
	), f"Temporal dimension {T} is not divisible by chunk_length {self.latent_chunk_duration}"
	return rearrange(latent, "b c (n t) h w -> (b n) c t h w", t=self.latent_chunk_duration)

Because of this, we can relax these asserts to only check for inputs that don't meet the minimum frame requirement. The error message is updated to clearly indicate the issue.

[pjannaty]

APPROVED_FOR_CI_RUN

[pjannaty]

pipelines/34098563

[aryamancodes]

Nice!