Research map / Image generation / Fast diffusion sampling
CoDA: From Text-to-Image Diffusion Models to Training-Free Dataset Distillation
Letian Zhou, Songhua Liu, Xinchao Wang
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 density-based core distribution discovery and alignment guidance | SDXL · ImageIDC IPC10 | accuracy: 40.1 -> 58.5 | Improved |
| CoDA vs prior SOTA (DiT+MGD3) on ImageIDC IPC50 | DiT+MGD3 · ImageIDC IPC50 | accuracy: 72.1 -> 77.6 | Improved |
| CoDA vs prior SOTA on ImageWoof at low IPC | MGD3 / Minimax · ImageWoof IPC10-20 (ResNet-18) | accuracy: IPC20: 42.5 (Minimax) vs 42.0 | Mixed |
| CoDA vs prior SOTA on ImageWoof at high IPC | MGD3 · ImageWoof IPC100 (ResNet-18) | accuracy: 68.8 -> 71.4 | Improved |
| CoDA vs IGD (ImageNet-trained model) at IPC10 | IGD · ImageNet-1K IPC10 | accuracy: 46.2 -> 44.3 | Hurt |
| CoDA vs IGD at IPC50 | IGD · ImageNet-1K IPC50 | accuracy: 60.3 -> 60.4 | Improved |
| CoDA vs D4M, an off-the-shelf competitor | D4M · ImageNet-1K IPC10 | accuracy: 27.9 -> 44.3 | Improved |
| apply CoDA to Places365 with no modification | SDXL · Places365 10 classes IPC10 | accuracy: 33.9 -> 47.0 | Improved |
| use K-Means on SDXL instead of nothing (ablation) | SDXL · Places365 10 classes IPC10 | accuracy: 33.9 -> 34.4 | No clear effect |
| add alignment stage (generated set G over real set R) (ablation) | CoDA discovery output · ImageIDC IPC50 | accuracy: 75.7 -> 77.6 | Improved |
| replace K-Means with CoDA discovery in UMAP-VAE space (ablation) | prototype selection · ImageIDC IPC10 | accuracy: 54.8 -> 58.5 | Improved |
| add UMAP preprocessing before clustering (ablation) | CoDA · ImageIDC IPC10 | accuracy: 45.8 -> 58.5 | Improved |
Closest papers by meaning
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.90)
- When Model Knowledge meets Diffusion Model: Diffusion-assisted Data-free Image Synthesis with Alignment of Domain and Class (ICML 2025 · similarity 0.89)
- FlashEval: Towards Fast and Accurate Evaluation of Text-to-image Diffusion Generative Models (CVPR 2024 · similarity 0.89)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.89)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Boosting Alignment for Post-Unlearning Text-to-Image Generative Models (NeurIPS 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.
