Team Ai
Datasetpublic

StonyBrook-CVLab/RareFace-50

RareFace-50 (from Low-Rank Head Avatar Personalization with Registers) Dataset for Low-Rank Head Avatar Personalization with Registers. Also available on arxiv. Project Page Dataset Summary RareFace-50 is a curated collection of challenging human faces intended for evaluating personalization of talking-head and avatar generation methods. Unlike many existing face video datasets that focus primarily on celebrities and well-known public figures (e.g., television… See the full description on the dataset page: https://huggingface.co/datasets/StonyBrook-CVLab/RareFace-50.

sourceHugging Facecc-by-nc-4.0updated 10mo agoView on Hugging Face
0likes41downloads
Dataset Card

RareFace-50 (from Low-Rank Head Avatar Personalization with Registers)

Dataset for Low-Rank Head Avatar Personalization with Registers. Also available on arxiv.

Project Page

Dataset Summary

RareFace-50 is a curated collection of challenging human faces intended for evaluating personalization of talking-head and avatar generation methods.

Unlike many existing face video datasets that focus primarily on celebrities and well-known public figures (e.g., television personalities), RareFace-50 deliberately targets underrepresented facial appearances, with an emphasis on:

  • —Distinctive facial details (e.g., pronounced wrinkles, unique tattoos, scars, or other high-frequency details),
  • —Wide variation in age and appearance,
  • —High-resolution, close-up footage.

The dataset is constructed from 50 identities, each with 2 short clips, for a total of 100 clips. Source videos are high-resolution interview-style recordings (1080p, 2K, and 4K) collected from YouTube public broadcasts. The average duration of each clip is around 15 seconds.

Important: This repository contains only metadata about the clips (YouTube links and temporal trim information) in a CSV file.

Dataset Structure

Files

The dataset is provided as a single CSV file in this repository (RareFace50.csv).

Each row corresponds to a two clip and includes:

  • —A YouTube link for the source video.
  • —Two start times and end times defining the clips within that video.
Timestamp format: All temporal fields are stored as strings in h:mm:ss format (e.g., 0:00:13, 0:01:05, 1:23:45).

Schema

  • —youtube_url (string) Full YouTube URL for the source video.
  • —start_time (string) Clip start time in h:mm:ss.
  • —end_time (string) Clip end time in h:mm:ss.

How to Use

Loading the CSV

You can access the CSV directly using Python’s standard tools or datasets:

python
from datasets import load_dataset

ds = load_dataset("StonyBrook-CVLab/RareFace-50")
print(ds["train"][0])

If you use this dataset, please be so kind to cite us:

@inproceedings{
chakkera2025lowrank,
title={Low-Rank Head Avatar Personalization with Registers},
author={Sai Tanmay Reddy Chakkera and Aggelina Chatziagapi and Md Moniruzzaman and Chen-ping Yu and Yi-Hsuan Tsai and Dimitris Samaras},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={(https://openreview.net/pdf?id=mhARf5VzCn)}
}