Research map / Image generation / Fast diffusion sampling
ScoreMix: Synthetic Data Generation by Score Composition in Diffusion Models Improves Recognition
Parsa Rahimi, Sébastien Marcel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ScoreMix synthetic augmentation to real data | IR50 ArcFace · IJB-C TAR@FPR 1e-6 | TAR: 70.37 -> 76.45 | Improved |
| add ScoreMix augmentation instead of scaling backbone to IR101 | IR50 ArcFace · IJB-B TAR@FPR 1e-6 | TAR: 34.84 -> 35.95 | Improved |
| use ScoreMix instead of AugGen augmentation | IR50 ArcFace · IJB-C TAR@FPR 1e-6 | TAR: 75.02 -> 76.45 | Improved |
| mix most distant classes in embedding space rather than closest (ablation) | ScoreMix · class selection average | Avg: 64.92 -> 67.44 | Improved |
| mix most distant classes in condition space (cosine) rather than closest (ablation) | ScoreMix · class selection average | Avg: 66.84 -> 66.95 | No clear effect |
| mix most distant classes in condition space (L2) rather than closest (ablation) | ScoreMix · class selection average | Avg: 65.23 -> 65.79 | No clear effect |
| choose embedding-distant pairs instead of random pairs (ablation) | ScoreMix · class selection average | Avg: 66.17 -> 67.44 | Improved |
| mix three classes instead of two (ablation) | ScoreMix · class selection average | Avg: 67.44 -> 65.69 | Hurt |
| select pairs by close condition and distant embedding (ablation) | ScoreMix · class selection average | Avg: 67.44 -> 66.98 | No clear effect |
| select pairs using a different discriminator (AdaFace-IR101) instead of self (ablation) | ArcFace-IR50 target · cross-architecture selection | Perf: 68.08 -> 67.62 | No clear effect |
| align generator outputs to class centers via regularization (ablation) | generator reproduction · reproduction-set recognition | Avg | Hurt |
| add ScoreMix augmentation to real data | IR50 ArcFace · training time on WebFace160K | hours: 2.54h -> 4.10h | Hurt |
Closest papers by meaning
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.87)
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (ICLR 2025 · similarity 0.86)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.85)
- DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024 · similarity 0.85)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.85)
- Neural Network-Based Score Estimation in Diffusion Models: Optimization and Generalization (ICLR 2024 · similarity 0.85)
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.
