Research map / Image generation / Fast diffusion sampling
Region-Adaptive Sampling for Diffusion Transformers
Ziming Liu, Yifan Yang, Chengruidong Zhang, Yiqi Zhang, Lili Qiu, Yang You, Yuqing Yang
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 |
|---|---|---|---|
| update only focused regions with RAS instead of uniformly fewer steps | Stable Diffusion 3 · COCO val2014, 4-5 steps | FID: 39.70 -> 31.99 at 1.43 -> 1.45 img/s | Improved |
| update only focused regions with RAS on Lumina-Next-T2I | Lumina-Next-T2I · COCO val2014, 5 steps | FID: 96.53 -> 53.93 | Improved |
| use RAS for end-to-end speedup | SD3 and Lumina-Next-T2I · 28/30 steps | speedup: 1x -> 2.44x (SD3), 1x -> 2.70x (Lumina) at 12.5% ratio | Improved |
| RAS on Lumina-Next-T2I human preference metrics at the same step count | Lumina-Next-T2I · ImageReward, 30 steps | ImageReward: 0.37 -> 0.13 | Hurt |
| human evaluation of RAS against dense inference | SD3 and Lumina-Next-T2I · human pairwise comparison | preference share: 28.29% dense vs 26.50% RAS vs 45.21% comparable | No clear effect |
| use dynamic sampling ratio instead of static frequency (ablation) | RAS · SD3 10 steps | FID: 37.92 -> 35.81 | Improved |
| choose tokens by region metric instead of randomly (ablation) | RAS · SD3 10 steps | FID: 43.19 -> 35.81 | Improved |
| add dense error-reset steps (ablation) | RAS · SD3 10 steps | FID: 46.10 -> 35.81 | Improved |
| cache keys and values of inactive tokens (ablation) | RAS · SD3 28 steps | FID: 31.36 -> 24.30 | Improved |
| cache keys and values of inactive tokens at 10 steps (ablation) | RAS · SD3 10 steps | FID: 32.33 -> 35.81 | Hurt |
| add starvation prevention (ablation) | RAS · SD3 10 and 14 steps | FID: 39.87 -> 35.81 (10 steps); 26.58 -> 26.48 (14 steps) | Mixed |
| add a second dense reset step (steps 7 and 11) (ablation) | RAS · SD3 14 steps | FID: 24.60 -> 24.58 | No clear effect |
Closest papers by meaning
- RAD: Region-Aware Diffusion Models for Image Inpainting (CVPR 2025 · similarity 0.90)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.90)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.89)
- Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (NeurIPS 2024 · similarity 0.89)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · 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.
