Optimize Your Sampling: Tuned Diffusion Sampling with Bayesian Optimization
OYS tunes diffusion sampling timesteps directly against the target quality metric.
The paper frames timestep selection as a black-box Bayesian optimization problem, rather than optimizing a proxy for sample quality. It reports gains over default schedules and Align Your Steps on text-to-image generation, plus improvements on inpainting and other image tasks. The method needs no extra training and can be used with distilled models and samplers including Euler and DPM-Solver++. A 5-step OYS schedule is reported to keep 89%-94% of 50-step quality while cutting inference cost by 10x. ArXiv · AI/CL/LG's note
The paper frames timestep selection as a black-box Bayesian optimization problem, rather than optimizing a proxy for sample quality. It reports gains over default schedules and Align Your Steps on text-to-image generation, plus improvements on inpainting and other image tasks. The method needs no extra training and can be used with distilled models and samplers including Euler and DPM-Solver++. A 5-step OYS schedule is reported to keep 89%-94% of 50-step quality while cutting inference cost by 10x. ArXiv · AI/CL/LG's note
score 5