Reproduce “UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition” on OpenResearch

About this paper

Xiao Lin, Yuge Huang, Jianqing Xu, Yuxi Mi, Shuigeng Zhou, Shouhong Ding

The UIFace framework from Tencent Youtu Lab and Fudan University enhances intra-class diversity in synthetic face images for Face Recognition (FR) models. It achieved an average accuracy of 93.27% on five benchmarks with 0.5M synthetic images, outperforming prior synthetic data methods and narrowing the performance gap with real-data trained models.

Key insights

• Identity-irrelevant attributes (e.g., expression, pose) are primarily formed during the early stages of the diffusion denoising process, while identity-relevant details are recovered in later stages.
• Generative models inherently possess the capability to produce diverse images, but this ability is suppressed when a fixed identity context is applied during conditional generation.
• Self-attention maps in diffusion models primarily influence identity-irrelevant attributes, whereas cross-attention maps are crucial for encoding identity-specific information.

Summary by alphaXiv.