Published June 5, 2026 | Version AVAPrintDB_v0
Dataset Open

AVAPrintDB: A Public Multi-Generator Avatar Fingerprinting Database and Benchmark

Authors/Creators

  • 1. ROR icon Universidad Autónoma de Madrid

Description

Description

AVAPrintDB is the first fully public multi-generator avatar fingerprinting resource, published in the paper "Leveraging Avatar Fingerprinting: A Multi-Generator Photorealistic Talking-Head Public Database and Benchmark", and combines:

  • A large-scale photorealistic talking-head avatar database,
  • Atandardized evaluation protocols,
  • Publicly available benchmark code,
  • Pretrained checkpoints,
  • Preprocessing pipelines,
  • and reproducibility resources.
The benchmark spans 66k+ avatar videos generated with three state-of-the-art avatar generators from distinct synthesis paradigms (GAGAvatar (Neurips 2024), LivePortrait (2025), and HunyuanPortrait (CVPR 2025)), two audiovisual source datasets (RAVDESS and CREMA-D), and multiple evaluation settings including generator shift, dataset shift, and demographic analysis.

Contents

Here we provide the database files used for the benchmark, containing:
  • MP4 files of photorealistic avatar videos, in videos.zip
  • Preprocessed landmarks for those videos, in landmarks.zip
  • Preprocessed embeddings from CLIP and DiNOv2, in embeddings.zip
  • Metadata for database expansion and reproducibility, in metadata.zip

For more details about the database generation process, the preprocessing steps and the associated benchmark code, please check the AVAPrintDB GitHub page.

File naming convention

For all the MP4 video files, preprocessed landmark files (*.npy) and preprocessed embeddings (*.pt), we follow the same naming convention:

<target_id>--<driving_id>--<UUID_of_driving_video>--<generator_acronym>.<ext>

Example:

Actor_09--Actor_16--aeca9daf-bdfd-553f-8287-92b94732bd61--GAGA.npy

Corresponds to a landmarks file with extension "npy", target_id = "Actor_09", driving_id = "Actor_16", UUID of driving video = "aeca9daf-bdfd-553f-8287-92b94732bd61", and generator used = "GAGA".

The mapping between UUIDs and the original driving video can be found in the metadata file avaprintdb_metadata.csv.

License

AVAPrintDB is released under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License, CC BY-NC-SA 4.0 

Citation

@article{pedrouzo2026leveraging,
  title={Leveraging Avatar Fingerprinting: A Multi-Generator Photorealistic Talking-Head Public Database and Benchmark},
  author={Pedrouzo-Rodriguez, Laura and Gomez, Luis F and Tolosana, Ruben and Vera-Rodriguez, Ruben and Daza, Roberto and Morales, Aythami and Fierrez, Julian},
  journal={arXiv preprint arXiv:2603.26934},
  year={2026}
}

Files

metadata.zip

Files (54.7 GB)

Name Size
md5:0f0f431ae6366ed75b14a8b85847e00b
31.0 GB Preview Download
md5:fa35742765a1e7f42d07abecb3320df7
7.6 GB Preview Download
md5:4864d678c9eafab248296716ac4d3b7c
1.1 MB Preview Download
md5:75fc83c6f948f5c29400ed04726ac886
16.1 GB Preview Download

Additional details

Related works

Is published in
Publication: 10.48550/arXiv.2603.26934 (DOI)
Is supplement to
Software: https://github.com/BiDAlab/AVAPrintDB (URL)

Software

Repository URL
https://github.com/BiDAlab/AVAPrintDB
Programming language
Python
Development Status
Active