Research map / Image generation / Fast diffusion sampling
Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models
Dang Nguyen, Jiping Li, Jinghao Zheng, Baharan Mirzasoleiman
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 |
|---|---|---|---|
| augment only slow-learnable examples with TADA synthetic images instead of USEFUL upsampling | ConvNeXt-T on CIFAR-10 (SGD) · CIFAR-10 | test error (%): 34.16 -> 27.40 | Improved |
| augment only slow-learnable examples with TADA instead of USEFUL upsampling | Swin-T on CIFAR-10 (SGD) · CIFAR-10 | test error (%): 14.93 -> 14.57 | Improved |
| add TADA to DiffuseMix when fine-tuning pretrained ResNet18 | DiffuseMix · Flowers-102 | test error (%): 8.92 -> 8.08 | Improved |
| add TADA to DiffuseMix when fine-tuning pretrained ResNet18 | DiffuseMix · Aircraft | test error (%): 25.65 -> 25.12 | Improved |
| augment all data with DiffuseMix versus no augmentation | pretrained ResNet18 · Flowers-102 | test error (%): 8.55 -> 8.92 | Hurt |
| augment only the 30-40% slow-learnable examples instead of all examples (Syn-all) or random subset (Syn-rand) (ablation) | TADA, ResNet18 · CIFAR-10 and CIFAR-100 | test error | Improved |
| increase amplification factor k above 2 with upsampling of real images (ablation) | USEFUL, ResNet18 CIFAR-10 · CIFAR-10 | test error | Hurt |
| increase amplification factor k to 5 with synthetic generation (ablation) | TADA, ResNet18 · CIFAR-10, CIFAR-100 | test error: best k=5 (k=4 on TinyImageNet) | Improved |
| generate synthetic images from noised real images instead of random noise (ablation) | TADA, ResNet18 · CIFAR-10 | test error | Improved |
| use 50 denoising steps instead of fewer or many more (ablation) | TADA with SGD and SAM · CIFAR-10 | test error | Mixed |
| select slow-learnable examples by clustering model outputs instead of high loss or misclassification (ablation) | TADA · appendix selection ablation | test error | Improved |
| apply TrivialAugment to targeted subset instead of full data (ablation) | upsampled data · CIFAR-10 | test error | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffuseMix — DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.89)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.89)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.88)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.88)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.87)
- Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification (CVPR 2025 · similarity 0.87)
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.
