Research map / Image generation / Text-to-image diffusion
Beyond Objects: Contextual Synthetic Data Generation for Fine-Grained Classification
William Yang, Xindi Wu, Zhiwei Deng, Esin Tureci, Olga Russakovsky
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 context preservation and context marginalization (BOB) to DataDream-style T2I fine-tuning | DataDream · Aircraft 5-shot | accuracy: 50.04 -> 57.37 | Improved |
| add context marginalization (sample background/pose across dataset) without preservation (ablation) | DataDream · Aircraft 10-shot, ImageNet ResNet-50 | accuracy: 68.00 -> 70.13 | Improved |
| add context preservation (attribute-enriched captions) without marginalization (ablation) | DataDream · Aircraft 10-shot, ImageNet ResNet-50 | accuracy: 68.00 -> 65.90 | Hurt |
| add context preservation on top of marginalization (ablation) | BOB · Aircraft 10-shot, ImageNet ResNet-50 | accuracy: 70.13 -> 73.78 | Improved |
| sample contexts across the whole dataset instead of only from the same class (ablation) | BOB · Aircraft 10-shot, ImageNet ResNet-50 | accuracy: 64.38 -> 73.78 | Improved |
| use all extracted dataset contexts instead of 100 GPT-generated backgrounds/poses (ablation) | BOB · Aircraft 10-shot, ImageNet ResNet-50 | accuracy: 72.10 -> 73.78 | Improved |
| use GPT-4o instead of Qwen2.5-VL-7B as captioning model (ablation) | BOB · Aircraft 10-shot, ImageNet ResNet-50 | accuracy | No clear effect |
| use weaker Qwen2.5-VL-3B captioning model (ablation) | BOB · Aircraft 10-shot, ImageNet ResNet-50 | accuracy: 70.88 | No clear effect |
| BOB instead of Diff-II on long-tail classification | Diff-II · CUB-LT IF=100 | accuracy: 56.10 -> 62.19 | Improved |
| BOB instead of Diff-II on easier long-tail benchmark | Diff-II · Flower-LT IF=50 and IF=10 | accuracy: 95.20 -> 95.68 (IF=50); 97.62 -> 96.13 (IF=10) | No clear effect |
| BOB instead of DataDream on Pets | DataDream · Pets 5-shot | accuracy: 92.67 -> 92.73 | No clear effect |
| augment 5 real images with BOB synthetic data versus 10 real images only | ImageNet ResNet-50 · Aircraft | accuracy: 55.48 -> 60.31 | Improved |
Closest papers by meaning
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.89)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.87)
- Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · similarity 0.87)
- DEFT: Decompositional Efficient Fine-Tuning for Text-to-Image Models (NeurIPS 2025 · similarity 0.87)
- Advancing Fine-Grained Classification by Structure and Subject Preserving Augmentation (NeurIPS 2024 · similarity 0.86)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.86)
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.
