Research map / Image generation / Editing and style transfer
UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UIFace, a diffusion-based synthetic face recognition framework that uses a learnable empty context with a two-stage sampling strategy and attention injection to raise intra-class diversity while preserving identity.
Problem
Synthetic face datasets for training face recognition models suffer from context overfitting under identity-conditioned diffusion, yielding low intra-class diversity and weaker FR accuracy than real data.
Approach
Train one diffusion model conditioned on either identity contexts or a learnable empty context; at inference, sample with the empty context in early timesteps (identity-irrelevant content) and switch to identity context in later timesteps, with an adaptive boundary from cross-attention differences and an attention injection that transfers empty-context attention maps into ID-conditioned generation
Key results
- UIFace with 0.5M synthetic images (10k IDs x 50) reaches 93.27% average verification accuracy over five benchmarks, above Arc2Face (91.73%) and CemiFace (92.30%).
- Scaling to 1.5M synthetic images (30k IDs x 50) gives 94.54% average, close to the CASIA-Real baseline of 95.05%.
- Ablation: the adaptive two-stage sampling plus attention injection raises ImRecall from 53.99 to 71.96 and average accuracy from 91.53 to 93.27 over the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UIFace instead of CemiFace with 0.5M synthetic images | synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 92.30 -> 93.27 | Helped |
| use UIFace with 1.0M synthetic images instead of Arc2Face with 1.2M | synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 93.14 -> 94.06 | Helped |
| add two-stage sampling with fixed partition (t0=500) (ablation) | baseline synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 91.53 -> 92.08 | Helped |
| replace fixed stage partition with adaptive partitioning (ablation) | two-stage synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 92.08 -> 92.70 | Helped |
| add attention injection (ablation) | baseline synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 91.53 -> 92.25 | Helped |
| combine fixed two-stage sampling with attention injection (ablation) | baseline synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 91.53 -> 92.71 | Helped |
| combine adaptive two-stage sampling with attention injection (ablation) | baseline synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 91.53 -> 93.27 | Helped |
| combine adaptive two-stage sampling with attention injection (ablation) | synthetic face dataset · Synthetic dataset | LPIPS: 0.5270 -> 0.5592 | Helped |
| combine adaptive two-stage sampling with attention injection (ablation) | synthetic face dataset · Synthetic dataset | Improved Recall: 53.99 -> 71.96 | Helped |
| use UIFace rather than baseline sampling | DDIM synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 91.53 -> 93.27 | Helped |
| use UIFace rather than baseline sampling | FPNDM synthetic face recognition · Average (five benchmarks) | face verification accuracy (%): 92.00 -> 93.59 | Helped |
Surprising findings
- Sampling with the empty context only in early timesteps increases diversity, and the authors report only a modest trade-off in intra-class identity similarity.
Limitations
- Gains over prior methods on some benchmarks are small (e.g., AgeDB and CALFW trail Arc2Face at 0.5M), and the method still does not fully match real-data training at equal scale.
- Adaptive partitioning relies on a threshold hyperparameter th and classifier-free guidance scale choices tuned on the experiments.
Experimental setup
Datasets: CASIA-WebFace
Benchmarks: LFW CFP-FP CPLFW AgeDB CALFW
Metrics: verification accuracy average accuracy LPIPS Improved Recall
Baselines: SynFace DigiFace DCFace IDiff-Face ID3 CemiFace Arc2Face
Builds on: Latent Diffusion Model classifier-free guidance DCFace IDiff-Face ArcFace
Closest papers by meaning
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.92)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.90)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.89)
- CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition (NeurIPS 2024 · similarity 0.89)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.88)
- Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.