Research map / Image generation / Fast diffusion sampling
Diagnosing and Improving Diffusion Models by Estimating Optimal Loss Value
Yixian Xu, Shengjie Luo, Liwei Wang, Di He, Chang Liu
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 |
|---|---|---|---|
| replace EDM2 schedule with optimal-loss-based schedule | EDM · CIFAR-10 conditional | FID: 1.94 -> 1.75 | Improved |
| optimal-loss-based training schedule | EDM · CIFAR-10 unconditional | FID: 1.98 -> 1.94 | Improved |
| optimal-loss-based training schedule | EDM · ImageNet-64 conditional | FID: 2.44 -> 2.25 | Improved |
| optimal-loss-based training schedule | Flow Matching (EDM sampler) · CIFAR-10 conditional | FID: 2.07 -> 1.77 | Improved |
| optimal-loss-based training schedule | Flow Matching (EDM sampler) · ImageNet-64 conditional | FID: 3.06 -> 2.29 | Improved |
| optimal-loss-based training schedule | LightningDiT · ImageNet-256 w/o CFG | FID: 2.29 -> 2.08 | Improved |
| optimal-loss-based training schedule | LightningDiT · ImageNet-256 w/ CFG | FID: 1.42 -> 1.30 | Improved |
| subtract optimal loss from training loss in scaling-law fit (ablation) | EDM2 120M-1.5B · ImageNet-64 large noise scale | correlation coefficient rho: 0.82 -> 0.94 | Improved |
| use uncorrected DOL (C=1) instead of corrected cDOL estimator (ablation) | optimal-loss estimator · CIFAR-10 | estimation error | Hurt |
| use epsilon-prediction target with large-sigma fit | NCSN (VE-eps) vs EDM · CIFAR-10 | FID / loss gap | Hurt |
Closest papers by meaning
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.91)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.91)
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.90)
- Evaluating the design space of diffusion-based generative models (NeurIPS 2024 · similarity 0.90)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.90)
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.90)
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.
