Research map / Image generation / Fast diffusion sampling
A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes USF, a unified exponential-integral sampling framework for diffusion ODE solvers, and S^3, a predictor-based search that finds per-timestep solver schedules for few-step training-free sampling of DPMs.
Problem
Training-free diffusion ODE samplers degrade sharply below 10 function evaluations; existing solvers fix solving strategies across timesteps, which is suboptimal for low NFE.
Approach
A unified exponential-integral framework splits each solver step into independent decisions (timestep, starting point, prediction type, Taylor order, derivative estimation, corrector usage), enabling per-timestep solver schedules that are searched automatically by a performance predictor in a multi-stage search.
Key results
- Achieves 2.69 FID on CIFAR-10 with 10 NFE and 6.86 FID with 5 NFE, versus 23.44 for the prior SOTA at 5 NFE.
- Applied to Stable Diffusion, achieves a 2x acceleration (10 NFE to 5 NFE) on MS-COCO 256x256 without sacrificing quality.
- Searched solver schedules outperform state-of-the-art training-free samplers on CIFAR-10, CelebA, ImageNet-64 and LSUN-Bedroom.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Search solver schedules with S³ at 5 NFE | Diffusion sampling · CIFAR-10 | FID: 23.44 -> 6.86 | Helped |
| Search solver schedules with S³ at 10 NFE | Diffusion sampling · CIFAR-10 | FID | Helped |
| Apply S³ to Stable Diffusion, reducing sampling from 10 NFE to 5 NFE | Stable Diffusion · MS-COCO 256×256 text-to-image generation | performance: from 10 NFE to 5 NFE | No clear effect |
Limitations
- Search requires a performance predictor and multi-stage search with nontrivial search cost
- Gains reported mainly for very low NFE settings
Experimental setup
Datasets: CIFAR-10 CelebA ImageNet-64 LSUN-Bedroom MS-COCO
Benchmarks: CIFAR-10 CelebA ImageNet-64 LSUN-Bedroom MS-COCO
Metrics: FID
Baselines: DPM-Solver DPM-Solver++ UniPC DEIS DDIM
Builds on: DPM-Solver DEIS UniPC DDIM
Closest papers by meaning
- Diffusion Sampling Correction via Approximately 10 Parameters (ICML 2025 · similarity 0.92)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.92)
- S4S: Solving for a Fast Diffusion Model Solver (ICML 2025 · similarity 0.91)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.91)
- Enhanced Diffusion Sampling via Extrapolation with Multiple ODE Solutions (ICLR 2025 · similarity 0.91)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.91)
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.