VOLBench / README.md
yang1232009's picture
Add shot beats and the seven-dimension evaluation package
aa7162d verified
|
Raw History Blame Contribute Delete
2.8 kB
metadata
pretty_name: VOLBench One-Pass Video Inputs
language:
  - en
  - zh
size_categories:
  - n<1K
tags:
  - text-to-video
  - video-generation
  - benchmark
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.jsonl

VOLBench One-Pass Video Inputs

VOLBench provides video generation inputs for 17 films and 384 shots. Each shot includes a whitebox reference video, numbered reference images, a generation prompt, and the initial storyboard beat. Campfire Square uses the original 16-shot version.

Dataset layout

  • Test split has one row per shot in film order. The ordinal field is one-based within each film. previous_shot_id is null at the start of a film.
  • Generation package contains the media and per-shot manifests. Paths in the test split are relative to the repository root. Reference files are shared by content hash, while each shot preserves its original picture order.
  • Evaluation package provides the seven-dimension rubric, beat manifests and scoring code. The dataset viewer shows the shot metadata and file paths. Download the repository snapshot to obtain the media and run generation or evaluation.

Download the evaluation code and beat manifests as a ZIP. Extract it into onepass_inputs/ so the evaluation/ folder sits beside films/ and manifest.json.

Generate and score

Use the whitebox as the video reference and upload the ordered images with the full generation prompt. Save the resulting clip as result.mp4 in the corresponding shot directory. Generate one result per shot.

After installing the evaluation requirements and setting EASYAPI_API_KEY, run these commands from the repository root:

python3 onepass_inputs/evaluation/run.py --check
python3 onepass_inputs/evaluation/run.py --preflight
python3 onepass_inputs/evaluation/run.py --out evaluation_results

The evaluation uses the current initial beat as the text target. The previous clip is used for the cut between shots. Reference frames used only for generation are excluded from scoring. This release contains generation inputs and evaluation code. Generated clips and evaluation scores are produced separately.

Prompt revision 4 retains 372 first-submission prompts and includes 12 skill-authored replacements. The original text of those 12 is kept beside the replacement. See prompt provenance.

Upload

After authenticating with the Hugging Face CLI, upload this folder to a dataset repository with hf upload OWNER/REPO . --repo-type dataset. Set the dataset's license and visibility according to the asset owner's release terms.