AIPeanutman/VidDub-Training-Data

VidDub Training Data

Upload in progress (2026-11-03): manifests are available; the 211.2 GiB media upload is still running. Check manifest/release.json before treating the package as complete.

The paper-selected synthetic H3 training media for VidDub, released as original-resolution source videos plus the exact clip-level selection used in the paper.

• Paper
• Code and preprocessing
• Project page
• Model weights
• VidDub-Bench

Dataset summary

• 45,221 synthetic source videos
• 85,435 selected training windows
• Source media: 1344×768, 24 FPS, approximately 4.46 seconds
• Model-ready clips: 81 frames, 512×512, 24 FPS
• Model-ready audio: 16 kHz mono
• Languages by source: 27,630 English / 17,591 Chinese
• Languages by selected window: 52,658 English / 32,777 Chinese
• Logical source-media size: approximately 211.2 GiB
• Frozen paper selection: manifest/release_selection.jsonl

The package deliberately stores original videos. The 512x512 face crops, aligned 16 kHz audio, masks, reference frames, and optional latent caches are derived locally by the source repository's data/ pipeline. This avoids distributing model-specific caches and allows users to inspect the original generated media.

Download

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

The full package requires at least 212 GiB before derived clips and caches. Download to a filesystem with sufficient free space.

Reproduce the processed clips

git clone https://github.com/LAW1223/VidDub.git
cd VidDub
pip install -r requirements.txt

export VIDDUB_DATA_ROOT=/path/to/VidDub-Training-Data
export VIDDUB_WEIGHT_ROOT=/path/to/weights
export VIDDUB_LATENTSYNC_ROOT=/path/to/LatentSync

python scripts/validate_training_dataset.py \
  --root "$VIDDUB_DATA_ROOT" \
  --manifests-only

python -m data.s01_analyze --shard 0 --num_shards 64
python -m data.s02_make_clips \
  --selection "$VIDDUB_DATA_ROOT/manifest/release_selection.jsonl" \
  --shard 0 --num_shards 64
python -m data.s03_cache_latents --shard 0 --num_shards 64

Launch every shard index from 0 through 63. s02_make_clips applies the frozen release selection at clip level: 40,214 sources contribute two accepted windows and 5,007 sources contribute one. Do not infer or materialize an unlisted sibling window.

Layout

videos/<shard>/<source_id>.mp4        original-resolution media
manifest/s00_manifest.jsonl           artifact-relative source manifest
manifest/release_selection.jsonl      exact selected clip IDs and crop metadata
manifest/s01_reference.jsonl          reference preprocessing measurements
manifest/quality_selected.jsonl       quality records for selected clips
manifest/release.json                 release contract and counts

manifest/s00_manifest.jsonl uses artifact-relative video_path values. release_selection.jsonl is authoritative for membership; dir is the relative destination of a derived clip, not a pre-generated file in this package.

Selection protocol

The frozen selection is the H3 portion of the paper's weights_v1 recipe. Candidates first require IQA ≥ 40 and absolute SyncNet offset < 4. The bottom 20% of the within-domain composite percentile of SyncNet confidence, IQA, and mouth sharpness is then removed. The published selection is already filtered; do not apply the quality rule again.

Data provenance and limitations

All videos are generated media with synthetic identities and are approved by the dataset owner for this release. Text prompts intentionally vary language, apparent age, gender presentation, ethnicity, setting, and emotion. These are generation instructions—not verified attributes of real people. Generated media may contain artifacts, stereotypes, demographic imbalance, inaccurate lip motion, or other model biases. The dataset contains synthetic depictions that include teenagers; it is not intended for identity recognition, biometric inference, or claims about real demographic groups.

License and responsible use

VidDub Training Data is released under CC BY-NC 4.0. Use it for lawful non-commercial research, education, and evaluation of consented video dubbing. Do not use it for impersonation, fraud, harassment, non-consensual sexual content, biometric surveillance, or deceptive political/media manipulation.

Report correction, removal, or licensing requests through the official VidDub repository.

Citation

@article{liu2026viddub,
  title   = {VidDub: Replacement-Aware Distillation for Two-Step Video Dubbing},
  author  = {Liu, Yexin and Lin, Yunlong and Wang, Zehan and He, Guangxin and
             Xu, Xinli and Zha, Jiajun and Liu, Yaofu and Wang, Yueze and
             Ping, Bowen and Li, Yuanming and Wang, Amie and Ghanem, Bernard and
             Li, Bing and Yang, Harry},
  journal = {arXiv preprint arXiv:2611.03404},
  year    = {2026}
}