Research map / Image generation / Text-to-image diffusion
IS SYNTHETIC DATA FROM GENERATIVE MODELS READY FOR IMAGE RECOGNITION?
Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, XIAOJUAN QI
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 synthetic data (LE+CF+SCE) to classifier tuning in zero-shot setting | CLIP-RN50 · 17 zero-shot datasets | top-1 accuracy (average): +4.31 | Improved |
| add synthetic data (LE+CF+SCE) to classifier tuning in zero-shot setting | CLIP-ViT-B/16 · 17 zero-shot datasets | top-1 accuracy (average): +2.90 | Improved |
| add synthetic data to zero-shot classifier tuning | CLIP-RN50 · EuroSAT zero-shot | top-1 accuracy: +17.86 | Improved |
| add language enhancement (LE) over basic label-name prompts (ablation) | zero-shot synthetic data · EuroSAT and CIFAR-100 zero-shot | top-1 accuracy: +6.71 on EuroSAT (CE); +0.66 on CIFAR-100 | Improved |
| add language enhancement (LE) over basic label-name prompts (ablation) | zero-shot synthetic data · CIFAR-10 zero-shot | top-1 accuracy: -0.68 (CE, LE vs B) | Hurt |
| add CLIP-based filtering (CF) to LE (ablation) | zero-shot synthetic data · CIFAR-10, CIFAR-100, EuroSAT zero-shot | top-1 accuracy: CIFAR-10 77.20 to 80.01 (CE) | Improved |
| soft cross-entropy (SCE) instead of CE loss (ablation) | zero-shot synthetic data · CIFAR-10, CIFAR-100, EuroSAT zero-shot | top-1 accuracy: +0.05 to +1.14 over CE (LE+CF: CIFAR-100 44.55 to 45.69) | Improved |
| train ResNet-50 from scratch on synthetic data only | ResNet-50 on CIFAR-100 · CIFAR-100 from scratch | top-1 accuracy | Hurt |
| tune more encoder parameters with synthetic data instead of classifier only (ablation) | CLIP-RN50 zero-shot on EuroSAT · EuroSAT zero-shot | top-1 accuracy: 55.37 (0.02%) down to 52.09 (92.19%) | Hurt |
| add synthetic data to few-shot training (basic strategy) (ablation) | CLIP-ViT-B/16 few-shot · EuroSAT 16-shot | accuracy: 87.1 vs 85.83 (no syn) | Improved |
| real guidance (RG) for synthetic generation vs basic (ablation) | few-shot synthetic data · EuroSAT 16-shot | accuracy: +1.37 over B | Improved |
| real filtering (RF) of synthetic images (ablation) | few-shot synthetic data · EuroSAT 16-shot | accuracy: +0.23 over B | Improved |
Closest papers by meaning
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.92)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.90)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.89)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.88)
- Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.88)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 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.
