Vec2Face Model Card

Project Page | Paper | Code | 🤗 Gradio demo

Introduction

Vec2Face is the first model that achieves the generated synthetic face recognition dataset (HSFace10K) first being higher than the same-scale real dataset (CASIA-WebFace).

Usage

You can directly download the model in this repository. You also can download the model in python script:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/6DRepNet_300W_LP_AFLW2000.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/arcface-r100-glint360k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/magface-r100-glint360k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/vec2face_generator.pth", local_dir="./")

For the image generation and training code, please go to Vec2Face github repository.

Performance

Datasets in 0.5M scale

This table compares the existing synthetic dataset generation methods on five standard face recognition test sets. The model trained with HSFace10K has better performance on CALFW than that trained with real dataset.

###Dataset Scaling up to 300K identities and 15M images

This is the uniqueness of the proposed Vec2Face, which can easily scale the dataset size up.

Other challenging test sets

We test the model performance on other four datasets, Hadrian (facial hair), Eclipse (face exposure), SLLFW (similar-looking), and DoppelVer (doppelganger).

Citation

@article{wu2024vec2face,
  title={Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors},
  author={Wu, Haiyu and Singh, Jaskirat and Tian, Sicong and Zheng, Liang and Bowyer, Kevin W},
  journal={arXiv preprint arXiv:2409.02979},
  year={2024}
}