Analyzing and Improving Fast Sampling of Text-to-Image Diffusion Models
This work addresses the challenge of efficient image generation for users of diffusion models, though it is incremental as it builds on existing training-free acceleration methods.
The paper tackles the problem of slow sampling in text-to-image diffusion models by proposing a new scheduling strategy called constant total rotation schedule (TORS), which outperforms previous training-free methods and achieves high-quality images with only 10 sampling steps on models like Flux.1-Dev and Stable Diffusion 3.5.
Text-to-image diffusion models have achieved unprecedented success but still struggle to produce high-quality results under limited sampling budgets. Existing training-free sampling acceleration methods are typically developed independently, leaving the overall performance and compatibility among these methods unexplored. In this paper, we bridge this gap by systematically elucidating the design space, and our comprehensive experiments identify the sampling time schedule as the most pivotal factor. Inspired by the geometric properties of diffusion models revealed through the Frenet-Serret formulas, we propose constant total rotation schedule (TORS), a scheduling strategy that ensures uniform geometric variation along the sampling trajectory. TORS outperforms previous training-free acceleration methods and produces high-quality images with 10 sampling steps on Flux.1-Dev and Stable Diffusion 3.5. Extensive experiments underscore the adaptability of our method to unseen models, hyperparameters, and downstream applications.