Fix: convert chunk_overlap from pixel to latent frames in autoregressive generation

[rubenohana]

Summary

generate_autoregressive_from_batch in video2world.py passes chunk_overlap directly as num_latent_conditional_frames to the model. However, chunk_overlap is in pixel frames while num_latent_conditional_frames expects latent frames. With the tokenizer's 4:1 causal temporal compression, this causes a mismatch between what the model conditions on and what pixel-space slicing assumes, producing artifacts at chunk boundaries during autoregressive video generation.

For example, with chunk_overlap=5:

• Before (bug): model conditions on 5 latent frames = 17 pixel frames, but only 5 pixel frames are skipped in concatenation
• After (fix): model conditions on get_latent_num_frames(5) = 2 latent frames = 5 pixel frames, matching the 5 pixel frames skipped

The fix converts chunk_overlap to latent frames only when passing to the model, and uses the original pixel value for all pixel-space operations. This matches the approach already used in cosmos-transfer2.5 for the equivalent chunked generation logic.

Why it went unnoticed

The default chunk_overlap=1, and get_latent_num_frames(1) = 1 — the conversion is identity for this value. The bug only manifests when chunk_overlap >= 2.

Changes

• Convert chunk_overlap (pixel) to latent frames via tokenizer.get_latent_num_frames() before passing to the model
• Keep all pixel-space operations (effective_chunk_size, chunk concatenation, input buffer update) using the original chunk_overlap value
• Update docstring to clarify that chunk_overlap and chunk_size are in pixel frames

Test plan

• Verify chunk_overlap=1 (default) produces identical output to the unfixed code (no behavior change for the common case)
• Verify chunk_overlap=5 produces smooth transitions at chunk boundaries (no artifacts)
• Verify output frame count is correct for various num_output_frames / chunk_overlap combinations

[aryamancodes]

Looks like you're failing lint. Can you run just lint?

[aryamancodes]

Thanks for the contribution and nice catch! MR LGTM!

[rubenohana]

No problems! I just ran ruff format to fix the lint issue. I think it should be good now!