Research map / Image generation / Fast diffusion sampling
Accelerating Diffusion Sampling with Optimized Time Steps
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.
Optimizes diffusion sampling time steps by minimizing a surrogate error bound for high-order ODE solvers (UniPC, DPM-Solver++), solved via constrained trust region in seconds, improving FID at 5 NFEs across pixel and latent DPMs.
Problem
Training-free diffusion ODE samplers with few steps use uniform (or hand-designed) time-step schedules, which are suboptimal when the number of sampling steps is small.
Approach
Derive an upper bound on the distance between the approximate ODE solution from a general local-polynomial explicit solver and the ground-truth x0, under a score-error assumption; minimize the bound over time steps with a constrained trust-region solver, yielding a plug-and-play, near-zero-cost optimized schedule.
Key results
- With UniPC and 5 NFEs, optimized steps reduce CIFAR-10 FID from 23.22 (uniform-λ) to 12.11.
- On ImageNet 256x256 latent DiT-XL-2 with 5 NFEs, FID improves from 23.48 to 8.66; on 512x512 from 20.28 to 11.40.
- Optimization runs in under 15 seconds for up to 15 NFEs on CPU (14.1 s at 15 NFEs).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize UniPC time steps instead of using uniform-λ steps | UniPC on CIFAR-10 32x32 · CIFAR-10 32x32, 5 NFEs | FID: 23.22 -> 12.11 | Helped |
| Optimize DPM-Solver++ time steps instead of using uniform-λ steps | DPM-Solver++ on CIFAR-10 32x32 · CIFAR-10 32x32, 5 NFEs | FID: 29.22 -> 12.91 | Helped |
| Optimize UniPC time steps instead of using uniform-λ steps | UniPC on ImageNet 64x64 · ImageNet 64x64, 5 NFEs | FID | Helped |
| Optimize UniPC time steps instead of using uniform-t steps | UniPC on ImageNet 256x256 · ImageNet 256x256, 5 NFEs | FID: 23.48 -> 8.66 | Helped |
| Optimize UniPC time steps instead of using uniform-t steps | UniPC on ImageNet 512x512 · ImageNet 512x512, 5 NFEs | FID: 20.28 -> 11.40 | Helped |
| Optimize UniPC time steps instead of baseline discretization schemes | UniPC on FFHQ 64x64 · FFHQ 64x64, 5 NFEs | FID | Helped |
| Optimize UniPC time steps instead of baseline discretization schemes | UniPC on AFHQv2 64x64 · AFHQv2 64x64, 5 NFEs | FID | Helped |
Surprising findings
- For UniPC on ImageNet 64x64, uniform-t is best at 5-6 NFEs while uniform-λ is best beyond 7, showing no single baseline schedule dominates.
Limitations
- Optimization objective is a surrogate bound under an assumed score error, not the exact sampling error; UniPC's corrector is not modeled.
- Choice of the power p in the error-scale function is set empirically (p=1 pixel, p=2 latent).
Experimental setup
Datasets: CIFAR-10 ImageNet 64x64 FFHQ 64x64 AFHQv2 64x64 ImageNet 256x256 ImageNet 512x512
Benchmarks: CIFAR-10 ImageNet 64x64 FFHQ 64x64 AFHQv2 64x64 ImageNet 256x256 ImageNet 512x512
Metrics: FID
Baselines: UniPC with uniform-t UniPC with uniform-λ UniPC with EDM schedule DPM-Solver++ with uniform-λ DPM-Solver++ with uniform-t DPM-Solver++ with EDM
Builds on: UniPC DPM-Solver++ EDM DDIM
Closest papers by meaning
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.92)
- Enhanced Diffusion Sampling via Extrapolation with Multiple ODE Solutions (ICLR 2025 · similarity 0.92)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.91)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.91)
- TADA: Improved Diffusion Sampling with Training-free Augmented DynAmics (NeurIPS 2025 · similarity 0.91)
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models (ICLR 2024 · 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.