VidDub-Bench

The 256-source evaluation benchmark introduced with VidDub. It contains 128 English and 128 Chinese synthetic talking-human clips, with fixed same-audio reconstruction and cross-audio replacement-dubbing conditions.

• Paper
• Code and benchmark runner
• Project page
• Model weights

Contract

• 256 source-disjoint clips
• 81 frames, 24 FPS, 512x512
• fixed source, mask, reference frame, paired audio, and replacement audio
• replacement audio follows the frozen cross_key assignment in the manifest
• manifest/items.jsonl is authoritative; row order and key values are immutable
• source clips are disjoint from the VidDub training and development splits

Download and validation

git clone https://github.com/LAW1223/VidDub.git
cd VidDub

huggingface-cli download AIPeanutman/VidDub-Bench \
  --repo-type dataset \
  --local-dir data/VidDub-Bench

python scripts/validate_benchmark.py --root data/VidDub-Bench

The validator checks every file hash, manifest field, fixed cross-audio assignment, rights record, and artifact-relative path. Add --media for a slower ffprobe check of all 256 video/audio pairs.

Layout

clips/<shard>/<key>/
  video.mp4
  audio.wav
  mask.png
  ref.png
manifest/
  items.jsonl
  rights.jsonl
  release.json
  checksums.sha256

Each item has a frozen language, seed, and cross_key. Use audio_same for reconstruction and audio_cross for replacement dubbing. All paths are relative to this package.

Run the released model

After following the model setup in the VidDub Quick Start:

python scripts/run_benchmark.py \
  --benchmark data/VidDub-Bench \
  --model weights/VidDub \
  --weight-root weights \
  --output outputs/VidDub-Bench \
  --condition both

The model is loaded once and reused across all selected rows. The runner writes same/<key>.mp4, cross/<key>.mp4, and a resumable results.jsonl receipt. See the full benchmark guide for smoke tests, partitioned execution, and reporting rules.

License and responsible use

VidDub-Bench is released under CC BY-NC 4.0. The clips use synthetic identities and are approved for this release; per-item records are in manifest/rights.jsonl.

Use is intended for non-commercial research and evaluation of consented video dubbing. Do not use this benchmark or derived outputs for impersonation, fraud, harassment, non-consensual sexual content, or deceptive political/media use.