Research map / Image generation / Editing and style transfer
HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere
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.
Generates synthetic face recognition datasets by packing identity embeddings on the ArcFace hypersphere via gradient-descent optimization with a gallery-based regularizer, then rendering images with an identity-conditioned diffusion face generator.
Problem
Synthetic face recognition datasets lack sufficient inter-class variation (identity separation), which limits their ability to replace real web-crawled face datasets for training face recognition from scratch.
Approach
Formulates identity generation as a packing (Tammes-style) problem on the face embedding hypersphere, maximizing minimum pairwise cosine distance between reference embeddings, with a regularizer pulling embeddings toward a gallery of generator-produced face embeddings to stay on the realistic embedding manifold; images are then generated by an identity-conditioned diffusion model with embedding an
Key results
- HyperFace with 10k identities and 500k images achieves 98.50 LFW, 84.23 CPLFW, 89.40 CALFW, 88.83 CFP, 86.53 AgeDB, competitive with DCFace (best on CALFW and AgeDB).
- Scaling HyperFace to 50k identities and 3.2M images reaches 85.60 CPLFW, 91.48 CALFW, 92.24 CFP, 90.40 AgeDB among the largest synthetic datasets, though DCFace-60k and Langevin-DisCo-30k remain stronger on some benchmarks.
- Using BUPT real images as the regularization gallery improves average accuracy over StyleGAN gallery (e.g., AgeDB 87.13 vs 87.07; CALFW 90.03 vs 89.82).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on HyperFace instead of DCFace | iResNet50 face recognition model · LFW | accuracy: 98.35 -> 98.50 | Helped |
| Train on HyperFace instead of DCFace | iResNet50 face recognition model · CPLFW | accuracy: 83.12 -> 84.23 | Helped |
| Train on HyperFace instead of DCFace | iResNet50 face recognition model · CALFW | accuracy: 91.70 -> 89.40 | Hurt |
| Train on HyperFace instead of DCFace | iResNet50 face recognition model · CFP-FP | accuracy: 88.43 -> 88.83 | Helped |
| Train on HyperFace instead of DCFace | iResNet50 face recognition model · AgeDB-30 | accuracy: 89.50 -> 86.53 | Hurt |
| Increase the number of identities from 10k to 50k (ablation) | HyperFace · CPLFW | accuracy: 84.68 -> 85.60 | Helped |
| Increase images per identity from 32 to 128 (ablation) | HyperFace · AgeDB | accuracy: 86.33 -> 87.62 | Helped |
| Increase gallery size from 10K to 50K (ablation) | HyperFace optimization · AgeDB | accuracy: 85.9 -> 86.85 | Helped |
| Use BUPT real images instead of StyleGAN images for the gallery (ablation) | HyperFace optimization · CALFW | accuracy: 89.82 -> 90.03 | Helped |
| Use LDM images instead of StyleGAN images for the gallery (ablation) | HyperFace optimization · LFW | accuracy: 98.67 -> 98.65 | No clear effect |
| Add regularization (alpha 0.50 rather than 0) (ablation) | HyperFace optimization · CALFW | accuracy: 88.87 -> 89.82 | Helped |
| Add embedding noise (beta 0.005 rather than 0) (ablation) | HyperFace image generation · AgeDB | accuracy: 85.90 -> 87.07 | Helped |
Surprising findings
- Random reference embeddings with a minimum cosine distance of 0.3 perform notably worse than HyperFace optimization, indicating that staying on the face embedding manifold matters more than raw spread.
- Removing training images most similar to FFHQ or CASIA-WebFace (identity leakage check) does not hurt recognition accuracy on benchmarks.
Limitations
- Full-batch optimization has quadratic complexity in number of identities, making scaling costly (84 hours for 50k identities).
- Regularization gallery and generator training data are still derived from real face images, so privacy concerns persist; possible demographic bias from pretrained components.
- Results on some benchmarks (e.g., CFP, AgeDB) fall behind DCFace and Langevin-DisCo at comparable scale.
Experimental setup
Datasets: HyperFace synthetic dataset (generated) StyleGAN-generated gallery BUPT CASIA-WebFace FFHQ LFW CALFW CPLFW CFP-FP AgeDB-30
Benchmarks: LFW CPLFW CALFW CFP-FP AgeDB-30
Metrics: verification accuracy (10-fold cross-validation)
Baselines: SynFace SFace Syn-Multi-PIE IDnet ExFaceGAN GANDiffFace Langevin-DisCo DigiFace-1M IDiff-Face DCFace
Builds on: DCFace ArcFace AdaFace StyleGAN
Closest papers by meaning
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.90)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.90)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.89)
- CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition (NeurIPS 2024 · similarity 0.89)
- Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion (CVPR 2025 · similarity 0.88)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 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.