Research map / Image generation / Fast diffusion sampling
Accelerating Guided Diffusion Sampling with Splitting Numerical Methods
Suttisak Wizadwongsa, Supasorn Suwajanakorn
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 |
|---|---|---|---|
| Split diffusion and conditioning into separate subproblems, using Strang splitting with PLMS4 for diffusion and PLMS1 for conditioning | Guided diffusion sampling · ImageNet256 | LPIPS: 0.111 ± .078 -> 0.072 ± .065 | Improved |
| Add Strang splitting instead of using Lie-Trotter splitting (ablation) | LTSP4 guided diffusion sampler · ImageNet128 | FID: 3.85 -> 3.78 | Improved |
| Add Strang splitting instead of using Lie-Trotter splitting (ablation) | Guided diffusion sampler · ImageNet256 | LPIPS: 0.111 ± .092 -> 0.072 ± .065 | Improved |
| Use PLMS4 for the diffusion subproblem while retaining PLMS1 for conditioning (ablation) | Lie-Trotter splitting · ImageNet256 | LPIPS | Improved |
| Use higher-order RK2 or RK4 instead of PLMS1 for the diffusion subproblem (ablation) | Lie-Trotter splitting · ImageNet256 | LPIPS at matched sampling time | Hurt |
| Use classical high-order methods for the conditioning subproblem (ablation) | Lie-Trotter splitting · ImageNet256 | LPIPS | Hurt |
| Use non-split PLMS4 rather than split STSP4 (ablation) | Guided diffusion sampler · ImageNet256, sampling time within 5 sec. | LPIPS: 0.072 ± .065 -> 0.240 ± .131 | Hurt |
| Use STSP4 rather than DDIM | Guided diffusion sampler · ImageNet256 | FID: 5.47 -> 4.49 | Improved |
| Use STSP4 rather than PLMS4 | Guided diffusion sampler · ImageNet512 | recall: 0.51 -> 0.45 | Hurt |
| Use STSP4 rather than DDIM | Guided diffusion sampler · Inpainting | LPIPS: 0.17 -> 0.16 | Improved |
| Use STSP4 rather than DDIM | Guided diffusion sampler · Colorization | LPIPS: 0.28 -> 0.26 | Improved |
| Use STSP4 rather than DDIM | Guided diffusion sampler · Super-resolution | LPIPS: 0.46 -> 0.42 | Improved |
Closest papers by meaning
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.91)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.90)
- Accelerating Diffusion Sampling with Optimized Time Steps (CVPR 2024 · similarity 0.90)
- Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models (ICML 2025 · similarity 0.89)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.89)
- TADA: Improved Diffusion Sampling with Training-free Augmented DynAmics (NeurIPS 2025 · 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.
