Research map / Image generation / Fast diffusion sampling
Fewer Denoising Steps or Cheaper Per-Step Inference: Towards Compute-Optimal Diffusion Model Deployment
Zhenbang Du, Yonggan Fu, Lifu Wang, Jiayi Qian, Xiao Luo, Yingyan Celine Lin
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 |
|---|---|---|---|
| add mixed-resolution denoising (low resolution in early steps) (ablation) | SD V1.5, 20 steps · MS-COCO 2014 val | FID: 18.42 -> 15.69 | Improved |
| add mixed-resolution denoising (ablation) | PixArt-alpha, 20 steps · MS-COCO 2014 val | FID: 29.16 -> 25.99 | Improved |
| add mixed-resolution denoising (ablation) | SDXL, 20 steps · MS-COCO 2014 val | FID: 14.10 -> 14.37 | No clear effect |
| add hybrid module caching on top of mixed-resolution (ablation) | SD V1.5, 20 steps · MS-COCO 2014 val | Clip Score: 30.78 -> 30.25 | Hurt |
| add hybrid module caching on top of mixed-resolution (ablation) | LCM, 8 steps · MS-COCO 2014 val | FID: 21.01 -> 22.76 | Hurt |
| reduce denoising steps from 20 to 8 | SD V1.5 · MS-COCO 2014 val | FID: 18.42 -> 20.60 | Hurt |
| reduce denoising steps from 20 to 8 | PixArt-alpha · MS-COCO 2014 val | FID: 29.16 -> 33.09 | Hurt |
| abandon CFG entirely (ablation) | SD V1.5 · MS-COCO 2014 val | FID: 18.42 -> 31.98 | Hurt |
| cache conditional (Cond) cross-attention instead of CFG-combined map (ablation) | DC+CA m=5 · MS-COCO 2014 val | FID: 103.71 -> 18.82 | Improved |
| increase cache interval k from 2 to 3 (ablation) | PostDiff · MS-COCO 2014 val | FID: 16.65 -> 26.96 | Hurt |
| abandon CFG earlier (m=0.45 instead of 0.75) (ablation) | PostDiff · MS-COCO 2014 val | Clip Score: 30.25 -> 28.82 | Hurt |
| raise resolution scale beta from 0.5 to 0.875 (ablation) | mixed-resolution denoising · MS-COCO 2014 val | FID | 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.
- DeepCache — DeepCache: Accelerating Diffusion Models for Free (CVPR 2024) · 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
- Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.93)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.92)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.92)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.91)
- Think while You Generate: Discrete Diffusion with Planned Denoising (ICLR 2025 · similarity 0.91)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · 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.
