EDMDPMSolverMultistepScheduler init_nose_sigma #10577
Description
When using the EDM DPM schedular, I have found that my results are qualitatively a lot better when starting sampling from unit Gaussian noise even though the sigma_max of my training distribution is 80. For reference, I am working on denoising trajectories not images, so better in this case just means the final sample is less noisy and smoother. Using the scheduler's builtin init_noise_sigma which is ~80 gives much worse results. I was wondering if anyone else has experienced this or could give any insight. Thanks.