Research map / Image generation / Fast diffusion sampling
Self-Refining Diffusion Samplers: Enabling Parallelization via Parareal Iterations
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 SRDS, which applies Parareal predictor-corrector iterations with a coarse DDIM solver and parallel fine DDIM block solves to speed up diffusion sampling without approximating the serial solution.
Problem
Diffusion model sampling requires hundreds of sequential model evaluations; prior speedups via fewer steps or distillation sacrifice sample quality, and prior parallel sampling (ParaDiGMS) has high memory and communication cost.
Approach
Use the Parareal algorithm from parallel-in-time ODE integration: a coarse 1-step DDIM sweep gives an initial trajectory, which is refined by independent fine √N-step DDIM block solves run in parallel followed by a sequential predictor-corrector sweep; the method provably converges to the serial solution in at most √N iterations.
Key results
- On pixel diffusion (N=1024), SRDS converges in 3.7-5.8 iterations with 147-212 effective serial evals (vs 1024 sequential) while keeping FID unchanged (e.g., LSUN Church 12.8 both).
- On StableDiffusion-v2 with 25 steps, pipelined SRDS reaches 1.7x wall-clock speedup vs serial, and 4.3x for 961 steps on 4 A100 GPUs, outperforming ParaDiGMS at thresholds 1e-3 and 1e-2.
- SRDS total model evaluations rise to about 5603 vs 1024 for LSUN Church, trading extra total compute for lower latency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SRDS to DDIM | Pretrained pixel diffusion models · LSUN Church, LSUN Bedroom, Imagenet, and CIFAR | FID: 12.8 -> 12.8; 10.0 -> 10.0; 9.0 -> 9.0; 7.6 -> 7.6 | No clear effect |
| Apply SRDS with maximum 1 iteration | StableDiffusion-v2 · COCO2017 captions, 100 steps | time per sample: 4.6 -> 2.0 | Helped |
| Apply SRDS with maximum 1 iteration | StableDiffusion-v2 · COCO2017 captions, 100 steps | CLIP score: 31.9 -> 31.9 | No clear effect |
| Apply SRDS with maximum 1 iteration | StableDiffusion-v2 · COCO2017 captions, 25 steps | time per sample: 1.2 -> 0.8 | Helped |
| Apply SRDS with maximum 1 iteration | StableDiffusion-v2 · COCO2017 captions, 25 steps | CLIP score: 31.7 -> 31.4 | Hurt |
| Increase SRDS maximum iterations from 1 to 3 (ablation) | StableDiffusion-v2 · COCO2017 captions, 25 steps | CLIP score: 31.4 -> 31.9 | Helped |
| Increase SRDS maximum iterations from 1 to 3 (ablation) | StableDiffusion-v2 · COCO2017 captions, 25 steps | time per sample: 0.8 -> 1.7 | Hurt |
| Add pipeline parallelism to SRDS (ablation) | StableDiffusion-v2 · 25 model evaluations | time per sample: 0.82 -> 0.69 | Helped |
| Add pipeline parallelism to SRDS (ablation) | StableDiffusion-v2 · 196 model evaluations | time per sample: 3.30 -> 2.85 | Helped |
| Add pipeline parallelism to SRDS (ablation) | StableDiffusion-v2 · 961 model evaluations | time per sample: 12.30 -> 10.31 | Helped |
| Use pipelined SRDS instead of ParaDiGMS at threshold 1e-1 | StableDiffusion-v2 · 25, 196, and 961 model evaluations | time per sample: 0.77 -> 0.69; 3.42 -> 2.85; 14.30 -> 10.31 | Helped |
| Increase tolerance τ from 0.1 to 0.5 (ablation) | SRDS on LSUN Church · LSUN Church | KID: 0.0146 -> 0.0147 | Hurt |
Surprising findings
- For 25-step StableDiffusion, vanilla SRDS with 3 iterations is slower than sequential (0.7x), so benefit depends on trajectory length and iterations.
- Early convergence occurs much sooner than the worst-case sqrt(N) bound, with longer trajectories converging faster in relative terms.
Experimental setup
Datasets: LSUN Church LSUN Bedroom ImageNet 64x64 CIFAR-10 COCO2017 captions DrawBench
Benchmarks: LSUN Church LSUN Bedroom ImageNet 64x64 CIFAR-10 COCO2017 captions (CLIP score) DrawBench
Metrics: FID KID CLIP score wall-clock time per sample effective serial evals total model evals
Baselines: ParaDiGMS ParaTAA Sequential DDIM
Builds on: Parareal DDIM ParaDiGMS
Closest papers by meaning
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.89)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.89)
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models (ICLR 2024 · similarity 0.89)
- S4S: Solving for a Fast Diffusion Model Solver (ICML 2025 · similarity 0.89)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.89)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.89)
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.