---
license: other
license_name: videvent-derived-research-only
license_link: https://arxiv.org/abs/2506.02448
pretty_name: ShotPilot Stage 1 Training Cache
task_categories:
- text-to-video
language:
- en
tags:
- video-generation
- multi-shot
- shot-transition
- wan2.1
- latents
size_categories:
- 1K<n<10K
viewer: false
---

# ShotPilot Stage 1 Training Cache

Precomputed training inputs for the ShotPilot Stage 1 causal multi-shot generator, built from the
[ShotPlan](https://huggingface.co/datasets/Pensioner/shotplan) training set: 6,404 clips, each 80 frames
(5 s at 16 fps) with one to three hard cuts. Videos were encoded at 480x832 with the
[Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B) VAE and prompts with its UMT5-XXL text encoder.
No raw video is included.

## Contents

| Folder | What it holds | Files | Size |
|---|---|---:|---:|
| `stage1_features_v0/` | Per-clip, per-shot VAE video latents | 6,414 | 26.7 GB |
| `stage1_memory_rgb6/` | Per-shot RGB memory: frames from earlier shots, VAE-encoded one at a time | 6,430 | 24.8 GB |
| `stage1_text_enriched_v2/` | UMT5-XXL encodings of the enriched global prompts | 6,409 | 12.0 GB |

Each folder has an `index.jsonl` (one line per clip, keyed by `metadata_row`) and a `manifest.json`.
Artifact paths in the index files are relative to the index file. Absolute paths inside manifests and
artifacts record where the cache was originally built and are not used for loading.

All tensors are `torch.bfloat16` `.pt` files (load with `torch.load(path, weights_only=True)`):

- **features** `rows/row_XXXXXX.pt`: `global_prompt`, `source_relative_path`, and `shots`, a list of
  `{sample_id, shot_index_in_window, first_chunk_action, latent_frames, latent}` with
  `latent` shaped `(16, T, 60, 104)`. Each shot was cropped to 4n+1 RGB frames before encoding.
- **memory** `shards/worker-XX/rows/row_XXXXXX.pt`: `memory_by_shot`, one entry per shot with
  `latent` `(16, K, 60, 104)`, `shot_ids` and `rgb_indices` (K is at most 6; the first shot has none).
- **text** `rows/row_XXXXXX.pt`: `prompt` and `context` `(tokens, 4096)` (83–474 tokens).
  `empty_context.pt` is the empty-prompt encoding used for text dropout and classifier-free guidance.

## Usage

```bash
hf download BlueSourceJY/shotpilot-stage1-data --repo-type dataset --local-dir data
```

This recreates `data/stage1_features_v0`, `data/stage1_memory_rgb6` and `data/stage1_text_enriched_v2`.
The ShotPilot loader checks SHA-256 digests of the index files and text rows, so keep the files byte-identical.
Train/validation/test splits are computed at load time from `source_relative_path` and a fixed seed.

## License and provenance

These caches are derived from ShotPlan, which is derived from [VidEvent](http://www.videvent.top)
(Liang et al., 2025). They inherit those terms: **non-commercial research use only**. If you use them,
cite ShotPlan and VidEvent and comply with VidEvent's usage terms.

```bibtex
@article{liang2025videvent,
  title={VidEvent: A Large Dataset for Understanding Dynamic Evolution of Events in Videos},
  author={Liang, Baoyu and Su, Qile and Zhu, Shoutai and Liang, Yuchen and Tong, Chao},
  journal={arXiv preprint arXiv:2506.02448},
  year={2025}
}
```

ShotPlan: <https://github.com/Pensioner-11/ShotPlan>