Research map / Image generation / Editing and style transfer
$\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for 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 ID^3, a conditional diffusion model with an ID-preserving loss and sampling algorithm plus a Tammes-based dataset generator, to produce synthetic face recognition training data that improves verification accuracy over prior SFR generators.
Problem
Synthetic face recognition with diffusion models lags far behind real-data training, due to poor identity preservation within classes and insufficient inter- and intra-class diversity in generated datasets.
Approach
Adds an inner-product identity term (face-recognition embedding alignment) to the diffusion ELBO-based loss, shown to equal maximizing a lower bound on an adjusted conditional likelihood; the sampler follows the score of this adjusted likelihood (original score plus scaled reversed-likelihood score), and a Tammes-problem anchor scheme generates maximally separated identities with perturbed intra-c
Key results
- ID^3 trained on FFHQ-based synthetic data reaches 86.50% average verification accuracy across five benchmarks, versus 84.09% for IDiff-Face.
- ID^3 trained on CASIA-WebFace reaches 89.80% average, slightly below DCFace's 89.56%? No: above it, but DCFace is better on LFW (98.55 vs 97.68) and CA-LFW (91.6 vs 90.73).
- Removing the reversed likelihood score drops average accuracy from 89.80% to 65.05% in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ID³ instead of IDiff-Face for synthetic data generation | LResNet50-IR trained on FFHQ-generated synthetic faces · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 84.09 -> 86.50 | Helped |
| use ID³ instead of IDiff-Face for synthetic data generation | LResNet50-IR trained on FFHQ-generated synthetic faces · AgeDB | face verification accuracy (%): 78.40 -> 83.78 | Helped |
| use ID³ instead of DCFace for synthetic data generation | LResNet50-IR trained on synthetic faces, CASIA setting · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 89.56 -> 89.80 | Helped |
| use ID³ instead of DCFace for synthetic data generation | LResNet50-IR trained on synthetic faces, CASIA setting · CA-LFW | face verification accuracy (%): 91.6 -> 90.73 | Hurt |
| add the reversed likelihood score to ID-preserving sampling (ablation) | ID³ · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 65.05 -> 89.80 | Helped |
| add attribute conditioning (ablation) | ID³ · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 88.27 -> 89.80 | Helped |
| use maximally separated identity anchors instead of random anchors (ablation) | ID³ dataset generation · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 85.59 -> 89.80 | Helped |
| use [0.7, 0.9] as the identity-similarity sampling bounds instead of the full method's bounds (ablation) | ID³ dataset generation · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 87.78 -> 89.80 | Hurt |
| use [0.5, 0.7] as the identity-similarity sampling bounds instead of the full method's bounds (ablation) | ID³ dataset generation · Average across LFW, CFP-FP, CP-LFW, AgeDB, and CA-LFW | face verification accuracy (%): 88.05 -> 89.80 | Hurt |
| increase intra-class pose variation in the generated dataset (ablation) | ID³ PoseOnly variants · CFP-FP | face verification accuracy (%) | Mixed |
| increase intra-class pose variation in the generated dataset (ablation) | ID³ PoseOnly variants · AgeDB | face verification accuracy (%) | No clear effect |
Surprising findings
- Removing the reversed-likelihood (identity) score collapses average accuracy to 65.05%, far below the variant with random anchors (85.59%).
Limitations
- Clear gap remains versus face recognition trained on real datasets such as MS1M.
- Attribute ranges limited to training distribution (age up to 70, pose within ±60 degrees), with weak extrapolation.
- DCFace uses real auxiliary images and still slightly outperforms ID^3 on some benchmarks.
Experimental setup
Datasets: FFHQ CASIA-WebFace
Benchmarks: LFW CFP-FP CPLFW AgeDB CALFW
Metrics: verification accuracy (%)
Baselines: IDiff-Face ID-Net DigiFace SFace SynFace DCFace
Builds on: denoising diffusion probabilistic models IDiff-Face DCFace pSp
Closest papers by meaning
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.92)
- Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion (CVPR 2025 · similarity 0.90)
- Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion (ICML 2025 · similarity 0.89)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.89)
- CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition (NeurIPS 2024 · similarity 0.88)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.88)
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.