Research map / Image generation / Fast diffusion sampling
DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease
Runsheng Bai, Chengyu Zhang, Yangdong Deng
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 |
|---|---|---|---|
| use DRiffusion aggressive mode with 4 devices | SD2.1/SDXL/SD3 · MS-COCO 2017 val | speedup: 3.6x (SD2.1, SDXL), 3.7x (SD3) | Improved |
| use DRiffusion conservative mode with 4 devices | SD2.1/SDXL/SD3 · MS-COCO 2017 val | speedup: 2.5x, 2.5x, 2.4x | Improved |
| apply DRiffusion parallel sampling | SD2.1 · MS-COCO 2017 val | FID: 23.63 -> 23.24 | No clear effect |
| apply DRiffusion aggressive mode, 4 devices | SD3 · MS-COCO 2017 val | HPSv2.1: 29.14 -> 27.64 | Hurt |
| apply DRiffusion aggressive mode, 4 devices | SDXL · MS-COCO 2017 val | PickScore: 22.43 -> 22.26 | Hurt |
| use DRiffusion instead of AsyncDiff at similar speedup | SDXL · MS-COCO 2017 val | PickScore: 22.02 -> 22.26 (3.4x vs 3.6x) | Improved |
| use DRiffusion instead of AsyncDiff | SDXL · MS-COCO 2017 val | extra memory: +554MB -> +222MB | Improved |
| use DRiffusion instead of halving steps to 25 | SDXL · MS-COCO 2017 val | FID: 24.51 -> 24.08 (2.0x vs 1.9x) | Improved |
| use conservative instead of aggressive mode (ablation) | DRiffusion · MS-COCO 2017 val | speedup vs HPSv2.1: SD2.1 4 dev: 3.6x/26.34 -> 2.5x/26.49 | Mixed |
Closest papers by meaning
- Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.91)
- AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising (NeurIPS 2024 · similarity 0.91)
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.91)
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.90)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Efficient Denoising Diffusion via Probabilistic Masking (ICML 2024 · similarity 0.90)
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.
