Research map / Image generation / Editing and style transfer
CemiFace: Center-based Semi-hard Synthetic Face Generation for 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 CemiFace, a conditional diffusion model conditioned on a similarity factor to the inquiry identity, generating center-based semi-hard synthetic faces for training face recognition models that beat DCFace on synthetic FR benchmarks.
Problem
Face recognition models trained on synthetic face datasets from existing generators suffer performance degradation because the generated samples lack sufficient discriminative quality; the paper asks which similarity-to-identity-center samples are most useful for training.
Approach
Empirically finds that training samples with mid-level similarity to their identity center (center-based semi-hard) work best, and builds a diffusion model conditioned on an identity embedding plus a scalar similarity factor m, with an auxiliary similarity-matching loss on the predicted clean image, so that generated faces have controllable similarity to the inquiry identity.
Key results
- At 0.5M data volume, CemiFace reaches 92.30 average verification accuracy versus 90.18 for reproduced DCFace, reducing the gap-to-real from 4.08 to 1.96.
- At 1.0M and 1.2M volumes, CemiFace achieves 93.07 and 93.22 average accuracy, halving the gap to the real CASIA-WebFace baseline (GtR 1.19 and 1.04).
- The optimal generation similarity m is around 0 (89.567 AVG on a 10k-identity ablation), and training with m sampled over [-1,1] gives the best results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace DCFace† with CemiFace for 0.5M synthetic data | Synthetic face recognition training · 0.5M synthetic data | AVG accuracy: 90.18 -> 92.30 | Helped |
| Replace DCFace† with CemiFace for 1.0M synthetic data | Synthetic face recognition training · 1.0M synthetic data | AVG accuracy: 91.44 -> 93.07 | Helped |
| Replace DCFace† with CemiFace for 1.2M synthetic data | Synthetic face recognition training · 1.2M synthetic data | AVG accuracy: 91.87 -> 93.22 | Helped |
| Train on middle-similarity images rather than the lowest-similarity group (ablation) | Face recognition model trained on CASIA-WebFace · CASIA-WebFace similarity groups | AVG accuracy: 82.36 -> 91.78 | Helped |
| Mix generation m over [0,0.1] instead of using fixed m 0 (ablation) | CemiFace · Generated data; training m [0,1] | AVG accuracy: 91.64 -> 91.11 | Hurt |
| Mix generation m over [-0.1,0.1] instead of using fixed m 0 (ablation) | CemiFace · Generated data; training m [0,1] | AVG accuracy: 91.64 -> 91.61 | No clear effect |
| Train with m range [-1,1] rather than [0,1] (ablation) | CemiFace · Generated data; generation m 0 | AVG accuracy: 91.64 -> 92.30 | Helped |
| Vary the training m interval from 0.04 to 0.02 or 0.06 while keeping range [-1,1] (ablation) | CemiFace · Generated data; generation m 0 | AVG accuracy: 92.28; 92.09; 92.30 | No clear effect |
| Add easy generation m 0.5 to semi-hard generation m 0 (ablation) | CemiFace · Generated data; training m [0,1] | AVG accuracy: 91.64 -> 90.36 | Hurt |
| Add easy generation m 0.9 or 1 to semi-hard generation m 0 (ablation) | CemiFace · Generated data; training m [0,1] | AVG accuracy: 91.64 -> 90.12; 91.64 -> 89.57 | Hurt |
| Use 1-shot Flickr rather than 1-shot WebFace inquiry images (ablation) | CemiFace trained with CASIA · 0.5M generated data | AVG accuracy: 91.64 -> 88.97 | Hurt |
| Use CosFace-pretrained rather than AdaFace-pretrained FR for the generator (ablation) | CemiFace · Generated data | AVG accuracy: 92.30 -> 92.60 | Helped |
Surprising findings
- The optimal generation similarity is around 0 for the synthetic generator, differing from the ~0.7 optimum found on real CASIA-WebFace data, which the authors attribute to the similarity model being pretrained on that dataset.
- Using a 1-shot Flickr inquiry set (challenging in-the-wild images) gives worse results, and a random center fails to converge.
Limitations
- Performance depends on the pretrained identity network used for similarity comparisons.
- Sensitive to inquiry image quality; blurred, occluded, or extreme-pose inquiries produce unclean samples.
- Relies on a pretrained face recognition model trained on data without user consent, raising privacy and legal concerns; still below the real-data CASIA-WebFace baseline.
Experimental setup
Datasets: CASIA-WebFace WebFace4M VGGFace2 Flickr LFW CFP-FP AgeDB-30 CALFW CPLFW DigiFace-1M
Benchmarks: LFW CFP-FP AgeDB-30 CALFW CPLFW
Metrics: 1:1 verification accuracy AVG accuracy Gap-to-Real FID
Baselines: DCFace DigiFace SynFace IDiff-Face CASIA-WebFace (real)
Builds on: DCFace DDPM DDIM AdaFace CosFace
Closest papers by meaning
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.89)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.89)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.88)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.88)
- Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion (CVPR 2025 · similarity 0.86)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.86)
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.