Research map / Image generation / Fast diffusion sampling
Parallel Sampling of Diffusion Models
Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, Nima Anari
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 |
|---|---|---|---|
| parallelize 100-step DDPM sampling with Picard iterations (ParaDiGMS) | DiffusionPolicy DDPM · Robosuite Square | time per episode: 37.0s -> 10.0s | Improved |
| parallelize 100-step DDPM sampling with ParaDiGMS | DiffusionPolicy DDPM · Robosuite Square | reward: 0.85 -> 0.85 | No clear effect |
| parallelize 100-step DDPM with ParaDiGMS | DiffusionPolicy DDPM · FrankaKitchen | time per episode: 112s -> 33.3s | Improved |
| parallelize 1000-step DDPM with ParaDiGMS on multiple A100 GPUs | StableDiffusion-v2 · COCO2017 captions | time per sample: 50.0s -> 14.6s | Improved |
| parallelize 1000-step DDPM with ParaDiGMS | StableDiffusion-v2 · COCO2017 captions | CLIP score: 32.1 -> 32.1 | No clear effect |
| parallelize 1000-step DDPM with ParaDiGMS | LSUN Church DDPM · LSUN Church | FID: 12.8 -> 12.9 | No clear effect |
| use ParaDDPM instead of 500-step DDIM for faster, higher-quality sampling | LSUN Church · LSUN Church | FID: 15.5 -> 12.9 | Improved |
| apply ParaDiGMS to 25-step DDIM (few steps) | DDIM 25-step · LSUN Church | speedup: 1.0x | No clear effect |
| increase tolerance tau from 1e-1 to 5e-1 (ablation) | ParaDDIM 200-step · StableDiffusion-v2 | CLIP score: 31.9 -> 31.5 | Hurt |
| increase tolerance tau from 1e-1 to 5e-1 (ablation) | ParaDDIM 200-step · StableDiffusion-v2 | speedup: 4.0x -> 4.9x | Improved |
| lower tolerance tau from 1e-1 to 5e-3 (ablation) | ParaDDIM 200-step · StableDiffusion-v2 | CLIP score: 31.9 -> 31.9 | No clear effect |
Closest papers by meaning
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.93)
- Accelerating Parallel Sampling of Diffusion Models (ICML 2024 · similarity 0.93)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.92)
- Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.92)
- AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising (NeurIPS 2024 · similarity 0.92)
- One Step Diffusion via Shortcut Models (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.
