Research map / Image generation / Fast diffusion sampling
When Higher Order Hurts: Pre-Asymptotic Order Collapse in Generative ODE Sampling — A Theory of Discretization–Learning Interaction
Farzad Salajegheh, Sudhir Mudur
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 |
|---|---|---|---|
| use Heun instead of Euler at low NFE (8) (ablation) | EDM · CIFAR-10 | FID-10k: 26.02 -> 91.91 | Hurt |
| use Heun instead of Euler at moderate NFE (16) (ablation) | EDM · CIFAR-10 | FID-10k: 12.03 -> 8.51 | Improved |
| use RK4 instead of Euler at low NFE (8) (ablation) | Rectified Flow · CIFAR-10 | FID-10k: 21.17 -> 42.05 | Hurt |
| use Heun at step T/8 (NFE 16) instead of RK4 at same step (NFE 32) (ablation) | Rectified Flow · CIFAR-10 | FID-10k: 12.92 -> 12.14 | Improved |
| add roughness-adaptive solver order selection | Rectified Flow · CIFAR-10 NFE 32 | FID-10k: 9.54 -> 8.01 | Improved |
| add roughness-adaptive solver order selection | EDM · CIFAR-10 NFE 8 | FID-10k: 26.02 -> 69.06 | Hurt |
| add roughness-adaptive solver order selection | EDM · CIFAR-10 NFE 16 | FID-10k: 12.03 -> 10.23 | Mixed |
| add roughness-adaptive solver order selection | Rectified Flow · CelebA-HQ NFE 16 | FID: 42.55 -> 35.25 | Improved |
| add roughness-adaptive solver order selection | Rectified Flow · CelebA-HQ NFE 64 | FID: 12.26 -> 12.68 | Hurt |
| add roughness-adaptive solver order selection | EDM · ImageNet NFE 16 | FID: 7.27 -> 11.41 | Hurt |
| use RK4 instead of Heun at high NFE (64) (ablation) | EDM · ImageNet NFE 64 | FID: 5.77 -> 5.76 | Improved |
Closest papers by meaning
- Learning to Solve Generative ODEs Beyond the Linear Span (NeurIPS 2026 · similarity 0.90)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.90)
- Bespoke Solvers for Generative Flow Models (ICLR 2024 · similarity 0.89)
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.89)
- SADA: Stability-guided Adaptive Diffusion Acceleration (ICML 2025 · similarity 0.88)
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models (ICLR 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.
