Research map / Image generation / Fast diffusion sampling
Fast Sampling Through The Reuse Of Attention Maps In Diffusion Models
Rosco Hunter, Łukasz Dudziak, Mohamed S. Abdelfattah, Abhinav Mehrotra, Sourav Bhattacharya, Hongkai Wen
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 perturbation search on top of HURRY (PHAST) (ablation) | HURRY, 10-step DPM++ with 3 reuse steps · ImageNet prompts | PSNR: 25.9 -> 27.2 | Improved |
| reuse attention maps instead of reducing steps to 13 | 20-step DDIM, SD1.5 · MS-COCO 2017 | PSNR: 17.50 -> 25.90 | Improved |
| reuse attention maps instead of reducing steps to 13 | SDXL · PartiPrompts | PSNR: 20.30 -> 29.92 | Improved |
| reuse attention maps instead of reducing steps to 13 | SD1.5 DDIM · MS-COCO 2017 | FID: 28.22 -> 29.18 | No clear effect |
| reuse attention maps instead of reducing steps to 13 | SD1.5 DDIM · MS-COCO 2014 | FID: 18.71 -> 17.62 | Improved |
| reuse attention maps instead of reducing steps to 13 | CFG-distilled SD1.5 DDIM · MS-COCO 2017 | PSNR: 17.68 -> 28.63 | Improved |
| store cached attention maps in FP16 (ablation) | PHAST, SD1.5 DDIM · MS-COCO 2017 | PSNR: 25.90 -> 25.95 | No clear effect |
| store cached attention maps in INT8 (ablation) | PHAST, CFG-distilled DPM++ · MS-COCO 2017 | PSNR: 25.34 -> 25.35 | No clear effect |
| reuse features (attention block outputs) instead of attention maps (ablation) | PHAST, DDIM, 10 reuse steps · ImageNet prompts | PSNR: 27.10 -> 25.05 | Hurt |
| reuse features instead of attention maps (ablation) | HURRY, DPM++, 10 reuse steps · ImageNet prompts | PSNR: 22.49 -> 20.21 | Hurt |
| use DPM++ instead of DDIM sampler | PHAST, SD1.5 · MS-COCO 2017 | PSNR: 25.90 -> 23.02 | Hurt |
Closest papers by meaning
- Grounded Text-to-Image Synthesis with Attention Refocusing (CVPR 2024 · similarity 0.91)
- Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects (CVPR 2025 · similarity 0.90)
- Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing (CVPR 2024 · similarity 0.90)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.89)
- Adapting Diffusion Models for Improved Prompt Compliance and Controllable Image Synthesis (NeurIPS 2024 · similarity 0.89)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 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.
