helensliang/AVEDIT-dataset

AVEDIT Dataset

A large multimodal dataset for audio-driven video generation/editing. The public release stores compact searchable indexes as Parquet and the original media/NumPy/JSON assets as uncompressed TAR shards that preserve the source-relative directory layout.

Main training samples

Image-augmented records are counted as independent training samples, as requested.

Split	Samples	Apparent size
person_plain	789,310	4.091 TB
person_with_image_aug	104,257	0.864 TB
object_sam_csv_task	111,711	0.330 TB
animal	88,015	0.477 TB
Total	1,093,293	5.763 TB

The four primary groups contain approximately 7.68 million physical files across source video/audio, masks, pose/landmark arrays, metadata, and augmented reference images.

Auxiliary data

• CloneTimbre/test500_0320: 1,626 reference records and a 500-record clone plan.
• CloneTimbre/_multinode_test: 29,377 reference triplets and 10,680 generated audio files.
• CloneTimbre/full_avatar3: partially generated working set with roughly 19k references.
• The source test/ and _ltx_dataset_check/ diagnostics are summarized locally but excluded from the public release because some metadata contains internal paths.

Storage format

• metadata/*.parquet: sample indexes and structured metadata. These are directly loadable with 🤗 Datasets.
• shards/<group>/part-xxxxx.tar: original assets, byte-preserving and stored with their original relative paths.
• tools/download_and_restore_avedit.py: download selected groups one shard at a time and restore the original directory tree.
• tools/restore_avedit_dataset.py: extract shards from an already downloaded snapshot.

The TAR files are intentionally not compressed because MP4, WAV, PNG, and NPY files gain little from compression, while uncompressed TAR is much faster to build, upload, download, and extract.

Load Parquet indexes directly

from datasets import load_dataset

plain = load_dataset(
    "helensliang/AVEDIT-dataset",
    "person_plain",
    split="train",
    streaming=True,
)
print(next(iter(plain)))

Download and restore original files

pip install -U huggingface_hub
python tools/download_and_restore_avedit.py \
  --output-dir /path/to/AVEDIT-dataset

Download only selected top-level groups:

python tools/download_and_restore_avedit.py \
  --output-dir /path/to/AVEDIT-dataset \
  --group person_plain \
  --group person_with_image_aug

Once extracted, the original relative paths referenced by JSON/CSV metadata are restored, so existing local-path training code can use the dataset without rewriting media records.

Completeness notes

• person_plain: 789,310 listed records; 789,276 source video/audio/mask files. Pose/bbox/landmark arrays exist for approximately 784.4k records.
• person_with_image_aug: 104,257 listed records; 104,255 source video/audio/mask files; 240,234 augmented reference PNGs. Pose/bbox/landmark arrays exist for 103,829 records.
• object_sam_csv_task: 111,711 rows; all derived audio/mask assets are present, while 110,484 source videos are listed/present.
• animal: 88,015 rows; all derived audio/mask assets are present, while 87,924 source videos are listed/present.

See metadata/file_counts.json, metadata/subset_summary.parquet, and metadata/summary.json for machine-readable details.

Public-release exclusions

The source tree occupies 6.041 TB. Approximately 52 GB of operational or internal-path artifacts are not published:

• hyvideo-avatar-3/: internal COS upload mapping caches containing infrastructure paths/URLs.
• _mmcopy_splits/: internal copy-job TSVs and logs.
• test/ and _ltx_dataset_check/ diagnostics containing internal paths.
• Raw CloneTimbre JSON metadata containing internal source paths/URLs; sanitized Parquet indexes are provided instead.
• Operational log directories and _scratch temporary outputs.

The expected public payload is approximately 5.988 TB, before small TAR-header overhead.

Responsible use

This dataset contains human faces, voices, speech transcripts, and derived biometric/pose information. Users are responsible for confirming that their use, redistribution, and generated outputs comply with applicable consent, privacy, copyright, personality-right, and data-protection requirements. No additional license to third-party source material is granted by this repository.