# CoAF 640x480 Training Data

This dataset contains the 640x480 CoAF v6 depth+track+RGB training data and preprocessed tensors.

Included archives:

- `full_video_latents.tar.zst`: precomputed 640x480 video latents.
- `full_image_latents.tar.zst`: precomputed condition image latents.
- `full_prompt_embeds.tar.zst`: precomputed prompt embeddings.
- `full_videos_and_images.tar.zst`: composed videos and condition images.
- `full_manifests_and_stats.tar.zst`: manifests and `track_norm_stats.pt`.
- `external_track_state_action.tar.zst`: track labels plus small raw state/action/instruction files referenced by manifests.
- `one_sample_overfit.tar.zst`: one-sample overfit dataset, including only its own latent/embed files.

Extract from the repository root that contains `coaf_dataset_24_25/`:

```bash
for f in *.tar.zst; do tar --zstd -xf "$f"; done
```

The current no-state0 v6 track training path does not require `state_norm_stats.pt` or `action_norm_stats.pt`. Track labels and `track_norm_stats.pt` are required.