Research map / Image generation / Fast diffusion sampling
IMS3: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation
Chenru Wang, Yunyi Chen, Zijun Yang, Joey Tianyi Zhou, Chi Zhang
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 Inversion-Matching fine-tuning (ablation) | DiT · ImageWoof IPC=10 | top-1 accuracy: 34.7 -> 37.3 | Improved |
| add Inversion-Matching fine-tuning (ablation) | DiT · ImageNette IPC=50 | top-1 accuracy: 82.9 -> 81.5 | Hurt |
| add Selective Subgroup Sampling (ablation) | DiT · ImageWoof IPC=10 | top-1 accuracy: 34.7 -> 40.9 | Improved |
| add Selective Subgroup Sampling (ablation) | DiT · ImageNette IPC=50 | top-1 accuracy: 82.9 -> 81.3 | Hurt |
| add both IM and S3 (ablation) | DiT · ImageWoof IPC=50 | top-1 accuracy: 49.3 -> 61.0 | Improved |
| add both IM and S3 (ablation) | DiT · ImageNette IPC=50 | top-1 accuracy: 82.9 -> 84.2 | Improved |
| use cosine 1-sigma loss instead of L1 in IM (ablation) | IMS3 · ImageWoof IPC=10 | top-1 accuracy: 38.7 -> 41.8 | Improved |
| use cosine 1-sigma loss instead of L2 in IM (ablation) | IMS3 · ImageNette IPC=50 | top-1 accuracy: 82.8 -> 84.2 | Improved |
| use ResNet-18 instead of CLIP as centroid feature extractor (ablation) | S3 · ImageWoof IPC=10 | top-1 accuracy: 38.1 -> 41.8 | Improved |
| use ResNet-18 instead of ResNet-50 as feature extractor (ablation) | S3 · ImageWoof IPC=10 | top-1 accuracy: 39.6 -> 41.8 | Improved |
| use real data instead of generated samples for centroids (ablation) | IMS3 · ImageWoof IPC=10 | top-1 accuracy: 39.6 -> 41.8 | Improved |
| increase lambda_IM to very large values (0.4-0.8) (ablation) | IM · ImageWoof | top-1 accuracy | Hurt |
Closest papers by meaning
- Taming Diffusion for Dataset Distillation with High Representativeness (ICML 2025 · similarity 0.91)
- MGD$^3$ : Mode-Guided Dataset Distillation using Diffusion Models (ICML 2025 · similarity 0.90)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Efficient Dataset Distillation via Minimax Diffusion (CVPR 2024 · similarity 0.89)
- Distribution-Aware Data Expansion with Diffusion Models (NeurIPS 2024 · similarity 0.88)
- D^4: Dataset Distillation via Disentangled Diffusion Model (CVPR 2024 · 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.
