Research map / Image generation / Fast diffusion sampling
Geometry-Aware Dataset Condensation for Diffusion Model Training
Xiao Cui, Yulei Qin, Mo Zhu, Wengang Zhou, Hongsheng Li, Houqiang Li
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 swap-based refinement (Stage II) after greedy selection (ablation) | geometry-aware selection · ImageNet 256x256 50K budget | FID-50K: 12.87 -> 11.01 | Improved |
| add swap-based refinement (Stage II) (ablation) | geometry-aware selection · ImageNet 256x256 10K budget | FID-50K: 3.82 -> 3.43 | Improved |
| add optimal transport loss L_OT (ablation) | geometry-aware selection · ImageNet 256x256 10K budget | FID-50K: 3.82 -> 3.43 | Improved |
| add mean-variance regularization L_sta (ablation) | geometry-aware selection · ImageNet 256x256 10K budget | FID-50K: 4.62 -> 3.43 | Improved |
| add confidence regularization L_conf (ablation) | geometry-aware selection · ImageNet 256x256 10K budget | IS: 337.0 -> 414.3 | Improved |
| use one-sided partial OT instead of balanced OT (ablation) | geometry-aware selection · ImageNet 256x256 10K budget | FID-50K: 3.54 -> 3.43 | Improved |
| increase capacity factor kappa from 1.05 to 1.10 (ablation) | one-sided partial OT · ImageNet 256x256 10K budget | FID-50K: 3.43 -> 3.58 | Hurt |
| vary dummy-source parameter gamma from 0.3 to 2.0 (ablation) | one-sided partial OT · ImageNet 256x256 10K budget | FID-50K: 3.43 -> 3.47 | No clear effect |
| replace difficulty-ranking selection (D2C) with OT-based distribution alignment | dataset condensation for DiT-L/2 · ImageNet 512x512 10K budget | FID: 14.8 -> 6.17 | Improved |
| replace D2C with geometry-aware selection on SiT-L/2 | dataset condensation for SiT-L/2 · ImageNet 256x256 100K budget | FID-50K: 15.01 -> 8.83 | Improved |
| replace D2C with geometry-aware selection, longer training | DiT-L/2 · ImageNet 256x256 50K budget, 300K iters | FID-50K: 4.76 -> 3.34 | Improved |
| geometry-aware selection vs D2C at 10K budget on SiT-L/2 (256x256) | SiT-L/2 · ImageNet 256x256 10K budget | FID-50K: 3.98 -> 3.56 | Improved |
Closest papers by meaning
- Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.90)
- Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation (NeurIPS 2025 · similarity 0.89)
- Distribution-Aware Data Expansion with Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Structure-Guided Adversarial Training of Diffusion Models (CVPR 2024 · similarity 0.87)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Composition and Alignment of Diffusion Models using Constrained Learning (NeurIPS 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.
