You need to agree to the conditions to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
AVID data is intended solely for the evaluation of deepfake detection systems and cannot be used for any commercial purpose.
Log in or Sign Up to review the conditions and access this dataset content.
AVID: Audio-Visual Identity-preserving Deepfakes (anonymous review copy)
Multilingual audio-visual deepfake benchmark.
(!) AVID data is intended solely for the evaluation of deepfake detection systems and cannot be used for any commercial purpose.
During peer review, this data is provided for the reviewers' assessment of the submission only. To protect reviewer anonymity, we do not record or use the requester information (HuggingFace username or email).
Contents
| Folder | Content |
|---|---|
fakes/ |
9,606 synthetic clips (tar shards) |
authentic/ |
25,203 authentic reference clips (tar shards) |
perturbed/<condition>/ |
Perturbation suite: 5 perturbation types x 3 severity levels, plus clean_none (tar shards) |
perturbed/noise_heavy/ |
Heavy-noise condition (tar shards) |
metadata.csv |
One row per fake/authentic clip |
34,809 clips in total, 162 subjects. The official evaluation split has 7,295 clips (3,580 authentic, 3,715 synthetic; 57 subjects), flagged by official_split=True. Evaluate only on this split. The perturbation suite covers it: 116,571 clips (109,276 perturbed + 7,295 clean); 149 heavy-noise fakes that could not be decoded are excluded.
metadata.csv columns
clip_id, path, subject, split (train/val/test), label (authentic/synthetic), language, official_split, video_model, voice_model, news_guided, source_clip, duration_sec, fps, width, height, md5, transcript. Generation columns are empty for authentic clips. A few fakes lack model/transcript fields.
Paths inside tar shards are relative: <Subject>/fakes/<video>/clip_XXX.mp4 or <Subject>/reference/clips/<video>/clip_XXX.mp4.
Intended use and ethics
- Subjects are deceased public figures only. Synthetic clips are generated for detection research and must not be used to impersonate, deceive or harass.
- Redistribution, commercial use and training of generative models are not permitted.
- Contains no personal data beyond publicly available footage of the subjects.
- Downloads last month
- 36