Research map / Image generation / Fast diffusion sampling
Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion
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 physics-inspired latent-space sampling (Langevin, Dispersion, DisCo) over a StyleGAN2 generator to build synthetic face recognition datasets, which train FR models that beat prior GAN-based datasets and approach diffusion-based ones.
Problem
Synthetic face datasets for training face recognition need identities that are sufficiently dissimilar (inter-class) and varied within identity (intra-class), and generative sampling alone yields insufficiently diverse identities; diffusion-based alternatives leak training data.
Approach
Treat latent samples as soft granular particles under repulsive contact forces, a Brownian (Langevin) noise term and a pull-back to the average latent, optimizing identity separation in a face-embedding metric; Dispersion adds intra-class variations via latent-space repulsion, and DisCo initializes these with StyleGAN covariate directions.
Key results
- Langevin-DisCo with 30k identities (1.95M images) reaches 98.97 LFW, 93.95 CALFW and 93.32 AgeDB, the best among synthetic datasets in the paper's table.
- Langevin-DisCo with 10k identities reaches 97.07 LFW and 83.38 AgeDB, outperforming GAN-based datasets such as SFace and IDnet on most benchmarks.
- DisCo's covariate-based initialization improves average accuracy over pure Dispersion (77.44 to 84.93 average with d0W=14, lambda0=1.5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| optimize identities with Langevin rather than random sampling (ablation) | Synthetic StyleGAN2 face dataset with Dispersion variations · LFW | accuracy: 90.78 -> 94.43 | Helped |
| increase Langevin iterations from 0 to 50 (d₀ᴱ=1.4) (ablation) | Synthetic face dataset with Dispersion variations · Average | accuracy: 73.38 -> 77.89 | Helped |
| use automatic time-step τ=0.3 instead of fixed δt=0.1 (ablation) | Langevin · Average | accuracy: 67.28 -> 67.39 | No clear effect |
| vary the stochastic force magnitude across η₀=0.003, 0.01, and 0.03 (ablation) | Langevin · Average | accuracy: 78.44; 78.02; 78.03 | No clear effect |
| increase embedding repulsion threshold from d₀ᴱ=1.2 to 1.5 (ablation) | Langevin · Average | accuracy: 72.97 -> 77.85 | Helped |
| increase embedding repulsion threshold from d₀ᴱ=1.5 to 1.6 (ablation) | Langevin · Average | accuracy: 77.85 -> 76.10 | Hurt |
| increase latent repulsion threshold from d₀ᵂ=8.0 to 14.0 (ablation) | Dispersion · Average | accuracy: 67.55 -> 80.69 | Helped |
| increase latent repulsion threshold from d₀ᵂ=14.0 to 20.0 (ablation) | Dispersion · Average | accuracy: 80.69 -> 52.19 | Hurt |
| initialize variation latents with covariate directions (DisCo, λ₀=1.5) instead of random-only initialization (ablation) | Dispersion · CPLFW | accuracy: 65.08 -> 75.85 | Helped |
| initialize variation latents with covariate directions (DisCo, λ₀=1.5) instead of random-only initialization (ablation) | Dispersion · CFP | accuracy: 65.6 -> 78.76 | Helped |
| increase covariate scale from λ₀=1.5 to 2.0 (at d₀ᵂ=16.0) (ablation) | DisCo · Average | accuracy: 84.93 -> 78.93 | Hurt |
| add training-set identity repulsion with dₜᵣᴱ=1.3 (ablation) | Langevin · Average | accuracy: 77.44 -> 78.38 | Helped |
Surprising findings
- Pure random sampling of identities yields FR models that are substantially worse than Langevin-optimized ensembles, with the gain continuing even after embedding distances plateau.
- Diffusion-based synthetic face datasets have been shown to memorize training data, motivating the GAN-based choice.
Limitations
- Pairwise interaction computation scales quadratically with number of identities.
- Still a gap between training on synthetic and real face recognition datasets.
- Pipeline depends on real data (StyleGAN training data and pretrained FR model), and results show sensitivity to hyperparameters such as d0 thresholds.
Experimental setup
Datasets: MS-Celeb-1M CMU Multi-PIE FFHQ IJB-C LFW CPLFW CALFW CFP-FP AgeDB-30
Benchmarks: LFW CPLFW CALFW CFP-FP AgeDB-30
Metrics: verification accuracy (10-fold)
Baselines: DigiFace-1M DCFace-0.5M DCFace-1.2M IDiff-Face SFace SFace2 Synface GANDiffFace IDnet ExFaceGAN Syn-Multi-PIE Random sampling (Reject)
Builds on: StyleGAN2 ArcFace AdaFace DCFace IDiff-Face
Closest papers by meaning
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.89)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.87)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.87)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.87)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.86)
- Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion (CVPR 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.