Region-Adaptive Sampling for Diffusion Transformers

Diffusion models (DMs) have achieved state-of-the-art performance across diverse generative tasks, yet their dependence on sequential forward passes fundamentally limits real-time efficiency. Existing acceleration methods primarily focus on reducing the number of sampling steps or reusing intermediate features. Leveraging the inherent flexibility of Diffusion Transformers (DiTs) in handling variable token counts, we introduce RAS, a training-free sampling strategy that dynamically adjusts update ratios across image regions based on model attention. Our key insight is that DiTs progressively focus on semantically meaningful regions, and such focused areas exhibit strong temporal continuity between consecutive steps. Building on this observation, RAS updates only these focused regions while reusing cached noise elsewhere, with focus maps inferred from the previous step's output. Experiments on Stable Diffusion 3 and Lumina-Next-T2I demonstrate up to 2.36× and 2.51× speedups, respectively, with negligible quality degradation-highlighting a practical pathway toward real-time diffusion transformer generation.