Diffusion models are among the best-performing generative models but are constrained by high computational costs. Previous work shows that optimizing timestep schedules with respect to factors such as the dataset, solver, and diffusion model can improve sampling efficiency and output quality. However, these schedules are static and do not adapt to individual diffusion processes. This work empirically shows that per-sample timestep schedules exist that can improve FID scores by up to 45.8% with an average of 31.4% over a globally optimal baseline on CIFAR-10. Computing such schedules efficiently is non-trivial due to the problem’s non-convexity and a tendency to converge to the global baseline. Despite this, up to a 12.1% FID reduction is achieved with a novel one-shot Timestep Prediction Model, while a recursive extension yields improvements of up to 18.2%. Both approaches were trained on theoretically optimal schedules and incur negligible computational overhead during sampling. This work proves that adaptive timesteps are a viable approach to further improve the sample quality and speed of diffusion models after other design choices have been made, such as distilling the diffusion model and selecting an ODE solver.