Research map / Image generation / Fast diffusion sampling
Otil: Accelerating Diffusion Model Inference via Communication-Efficient Multi-GPU Parallelism
Xin Li, Shujun Tian, Tao Lu, Han Bao, Zonghui Wang, Wenzhi Chen
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 |
|---|---|---|---|
| transmit only 1/4 of latent sub-blocks instead of all (ablation) | Otil · SDXL 1.0, 2 GPUs | per-step latency (ms): 43.5324 -> 27.5324 (LPIPS 0.0139 -> 0.0142) | Improved |
| transmit only 1/16 of sub-blocks (ablation) | Otil · SDXL 1.0, 2 GPUs | LPIPS: 0.0142 -> 0.1342 | Hurt |
| transmit only 1/4 sub-blocks instead of full on 4 GPUs (ablation) | Otil · SDXL 1.0, 4 GPUs | LPIPS: 0.0693 -> 0.1350 | Hurt |
| use 8x8 sub-blocks instead of 4x4 (ablation) | Otil · SDXL 1.0, 2 GPUs | LPIPS: 0.0571 -> 0.0142 | Improved |
| use 8x8 sub-blocks instead of 16x16 (ablation) | Otil · SDXL 1.0, 2 GPUs | LPIPS: 0.0294 -> 0.0142 | Improved |
| rank sub-blocks by cosine similarity instead of random, SSIM, mutual information or dHash (ablation) | Otil · latent-to-pixel change alignment (AUC) | AUC | Improved |
| add dynamic polling so every sub-block is revisited cyclically | Otil · COCO Captions 2014 | image quality | Improved |
| combine with few-step samplers (UniPC) and 2 GPUs | Otil · SDXL 1.0, 30 steps | speedup: 1.66 -> 2.84 | Improved |
| combine with LoRA SD 1.5 on 2 GPUs | Otil · SD 1.5 LoRA, 30 steps | speedup: 1.78 -> 2.46 | Improved |
| Otil instead of DistriFusion | SD 1.5, 2 GPUs · COCO Captions 2014 | LPIPS: 0.041 -> 0.052 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PipeFusion — PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models (CVPR 2024 · similarity 0.92)
- AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising (NeurIPS 2024 · similarity 0.92)
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.92)
- Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.92)
- Accelerating Parallel Diffusion Model Serving with Residual Compression (NeurIPS 2025 · similarity 0.90)
- PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference (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.
