Research map / Image generation / Fast diffusion sampling
Diffusion-Based Data Augmentation for Image Recognition: A Systematic Analysis and Evaluation
Zekun Li, Yinghuan Shi, Yang Gao, Dong Xu
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 |
|---|---|---|---|
| use Full Concatenation instead of Full Replacement for synthetic samples (ablation) | GIF · Caltech-101 (from scratch) | top-1 accuracy: 51.43 -> 76.11 | Improved |
| use Full Concatenation instead of Full Replacement (ablation) | GIF · CIFAR-100 (from scratch) | top-1 accuracy: 31.78 -> 68.23 | Improved |
| use Local Random Replacement instead of Full Concatenation with pretrained classifier (ablation) | Diff-II · Birds 10-shot (pretrained) | top-1 accuracy: 71.07 -> 72.96 | Improved |
| use Full Replacement instead of Local Random Replacement (ablation) | Diff-II · Birds 5-shot (pretrained) | top-1 accuracy: 61.29 -> 41.07 | Hurt |
| use Full Replacement instead of Full Concatenation (structure-preserving generation) (ablation) | DiffuseMix · Birds 5-shot (pretrained) | top-1 accuracy: 53.72 -> 55.88 | Improved |
| use Global instead of Local Random Replacement (ablation) | Diff-II · Birds 10-shot (pretrained) | top-1 accuracy: 72.96 -> 72.47 | No clear effect |
| raise SDEdit transition strength from 0.1 to 0.9 (ablation) | Real Guidance · Caltech-101 (from scratch) | top-1 accuracy: 54.96 -> 65.38 | Improved |
| raise SDEdit transition strength from 0.1 to 0.9 (ablation) | Real Guidance · ImageNet-100 (from scratch) | top-1 accuracy: 66.30 -> 77.12 | Improved |
| add DreamBooth-LoRA UNet fine-tuning on top of Textual Inversion (ablation) | DA-Fusion (vs Diff-Aug) · Aircraft few-shot average | top-1 accuracy: 32.53 -> 34.36 | Improved |
| add inter-class image mixup in generation (ablation) | Diff-Aug · Birds few-shot average | top-1 accuracy: 50.11 -> 53.67 | Improved |
| fine-tune diffusion model (Diff-Mix) instead of untuned low-strength generation on medical data | Real Guidance · Blood 25/class (from scratch) | top-1 accuracy: 86.67 -> 70.44 | Hurt |
| apply DiffDA at fine species granularity | Diff-Mix · Semi-iNat species level | top-1 accuracy: 41.88 -> 37.04 | Hurt |
Closest papers by meaning
- Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification (CVPR 2025 · similarity 0.89)
- Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · similarity 0.88)
- DiffAug: A Diffuse-and-Denoise Augmentation for Training Robust Classifiers (NeurIPS 2024 · similarity 0.88)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.88)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- Towards Universal Dataset Distillation via Task-Driven Diffusion (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.
