Region-Adaptive Sampling for Diffusion Transformers

The article presents Region-Adaptive Sampling (RAS), a novel training-free sampling strategy for Diffusion Transformers (DiTs) that optimizes real-time performance by dynamically assigning different sampling ratios to image regions based on the model's focus. RAS leverages temporal consistency by updating only the semantically meaningful areas, utilizing cached noise for less relevant regions, resulting in speedups of up to 2.36x on Stable Diffusion 3 and 2.51x on Lumina-Next-T2I with minimal quality loss. This advancement is crucial for practitioners aiming to enhance the efficiency of diffusion models in real-time applications.