Research map / Image generation / Fast diffusion sampling
AugGen: Synthetic Augmentation using Diffusion Models Can Improve 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 AugGen, a self-contained synthetic augmentation method that trains a class-conditional diffusion model on the target face dataset and samples mixed-identity classes via an alpha/beta grid search to augment real training data for face recognition.
Problem
Face recognition training with limited real data and privacy concerns; existing synthetic face data methods rely on external datasets or auxiliary/pretrained models.
Approach
Train a class-conditional diffusion generator and a discriminator only on the same real FR dataset, then synthesize new identities by linearly mixing two class conditions (alpha*c_i + beta*c_j), selecting alpha/beta by a dissimilarity-plus-similarity score computed with the discriminator's features, and augment real data with these samples.
Key results
- On CASIA-WebFace, AugGen mixed with real data improves IJB-C TAR@FPR=1e-5 to 18.58% vs 5.37% for CASIA-WebFace alone (IR50).
- On WebFace160K, adding 600K AugGen samples raises IJB-B TAR@FPR=1e-6 from 32.13 to 36.62 and IJB-C 1e-5 from 78.81 to 85.02.
- Adding ~600K AugGen samples yields gains comparable to adding ~110K real images (Ratio 1.69), per the paper's additional-real-data experiment.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 0.6M AugGen samples to 0.5M CASIA-WebFace images | IR50 trained on CASIA-WebFace · IJB-B/C and TinyFace | TAR or rank-1 accuracy: 1.02±0.26 -> 2.61±0.91; 5.06±1.70 -> 15.74±3.20; 0.73±0.19 -> 4.36±1.41; 5.37±1.41 -> 18.58±3.99; 58.12±0.31 -> 59.82±0. | Helped |
| Add 0.6M AugGen samples to 0.5M CASIA-WebFace images | IR50 trained on CASIA-WebFace · Avg-H | accuracy: 94.21±0.09 -> 94.66±0.03 | Helped |
| Add 0.6M AugGen samples to 0.16M WebFace160K images | IR50 trained on WebFace160K · IJB-B/C TAR at FPR 1e-6 and 1e-5 | TAR: 32.13±1.87 -> 36.62±0.77; 72.18±0.18 -> 78.32±0.33; 70.37±0.75 -> 78.58±0.15; 78.81±0.32 -> 85.02±0.15 | Helped |
| Add 0.6M AugGen samples to 0.16M WebFace160K images | IR50 trained on WebFace160K · TinyFace rank-1 | rank-1 accuracy: 61.51±0.16 -> 61.60±0.38 | No clear effect |
| Add 0.6M AugGen samples to 0.16M WebFace160K images | IR50 trained on WebFace160K · Avg-H | accuracy: 92.50±0.02 -> 94.17±0.08 | Helped |
| Use IR101 instead of IR50, without synthetic augmentation | CASIA-WebFace real-data training · IJB-B/C and TinyFace | TAR or rank-1 accuracy: 1.02±0.26 -> 0.74±0.31; 5.06±1.70 -> 3.94±1.62; 0.73±0.19 -> 0.38±0.13; 5.37±1.41 -> 3.92±1.96; 58.12±0.31 -> 59.64±0.49 | Mixed |
| Add AugGen samples rather than use the real-data IR101 upgrade | CASIA-WebFace-trained recognition model · TinyFace rank-1 | rank-1 accuracy: 59.64±0.49 -> 59.82±0.13 | Helped |
| Add 600K AugGen samples instead of adding 110K real images | WebFace160K-trained IR50 · IJB-B/C TAR and Avg-H | TAR or accuracy: 36.62{ ±0.77} vs 33.53{ ±1.47}; 78.32{ ±0.33} vs 78.26{ ±0.05}; 78.58{ ±0.15} vs 78.49{ ±0.54}; 85.02{ ±0.15} vs 85.02{ | Mixed |
| Increase AugGen synthetic samples from 200K to 600K | WebFace160K-trained IR50 · Avg-H | accuracy: 93.77{ ±0.04} -> 94.17{ ±0.08} | Helped |
Surprising findings
- Mixing AugGen data with the CASIA-WebFace real data performs better than training on real data alone, but some other synthetic methods (e.g., DCFace) degrade performance on IJB-B/C when combined with CASIA-WebFace.
- The paper reports that current generative metrics such as FD and KD correlate poorly with downstream discriminative performance.
Limitations
- Computational cost of training the generator from scratch on each target dataset.
- Results are shown only in a controlled setting with generators trained from scratch; pre-trained generators are not explored.
- Margin losses in FR are not compatible with soft labels, so mixed classes are treated as new hard classes.
Experimental setup
Datasets: CASIA-WebFace WebFace160K (subset of WebFace4M) WebFace4M
Benchmarks: LFW CFP-FP CPLFW CALFW AgeDB IJB-B IJB-C TinyFace
Metrics: TAR@FPR=1e-6 TAR@FPR=1e-5 rank-1 accuracy verification accuracy
Baselines: DigiFace1M RealDigiFace IDiff-face DCFace CASIA-WebFace IR101 WebFace160K IR101
Builds on: latent diffusion AdaFace ArcFace DCFace IDiff-face
Closest papers by meaning
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.88)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.88)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.88)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.87)
- UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation (NeurIPS 2025 · similarity 0.87)
- Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion (ICML 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.