This paper proposes an optimization-based method to accelerate diffusion sampling by finding appropriate time steps. Diffusion probabilistic models (DPMs) have shown strong performance in image generation but require many sampling steps, which is inefficient. The authors introduce a framework to optimize time steps for numerical ODE solvers, aiming to minimize the distance between the true solution and the approximate solution. This optimization is efficiently solved using the constrained trust region method, taking less than 15 seconds. The method is tested on both pixel and latent space DPMs with unconditional and conditional sampling. When combined with the state-of-the-art UniPC solver, the optimized time steps significantly improve image generation performance, as measured by FID scores, on datasets like CIFAR-10 and ImageNet. The results show that the optimized time steps lead to lower FID scores, indicating better image quality. The method is also efficient, with a negligible computational cost compared to learning-based approaches. The paper demonstrates that the proposed optimization improves sampling efficiency and quality, making diffusion models more practical for real-world applications.This paper proposes an optimization-based method to accelerate diffusion sampling by finding appropriate time steps. Diffusion probabilistic models (DPMs) have shown strong performance in image generation but require many sampling steps, which is inefficient. The authors introduce a framework to optimize time steps for numerical ODE solvers, aiming to minimize the distance between the true solution and the approximate solution. This optimization is efficiently solved using the constrained trust region method, taking less than 15 seconds. The method is tested on both pixel and latent space DPMs with unconditional and conditional sampling. When combined with the state-of-the-art UniPC solver, the optimized time steps significantly improve image generation performance, as measured by FID scores, on datasets like CIFAR-10 and ImageNet. The results show that the optimized time steps lead to lower FID scores, indicating better image quality. The method is also efficient, with a negligible computational cost compared to learning-based approaches. The paper demonstrates that the proposed optimization improves sampling efficiency and quality, making diffusion models more practical for real-world applications.