Noisy results with "order == 1" (trying to replicate DDIM resutls) about dpm-solver

[LuChengTHU]

Hi @WikiChao , does your code contain this line?: https://github.com/LuChengTHU/dpm-solver/blob/main/dpm_solver_pytorch.py#L105

If so, could you please print the first 5 and last 5 items of log_alphas?

[WikiChao]

Thanks for the prompt reply! The trick did help, it seems I am using the previous version and missing such a line of code.

The results make sense now, but they are still worse than DDIM. I have tried different settings, e.g., "multistep" or "single step", "order = 2 or 3", "step = 10 to 100", but cannot beat DDIM. Are there any tricks in choosing hyperparameters, for example, clipping log-SNR by different values?

Thanks a lot!

Chao

[LuChengTHU]

Hi @WikiChao ,

"but they are still worse than DDIM": In fact, order=1 is exactly the DDIM. You can try to reproduce the results of DDIM by manually setting the timestep in

dpm-solver/dpm_solver_pytorch.py

Line 453 in 5c6ee9f

def get_time_steps(self, skip_type, t_T, t_0, N, device):

as the same as your DDIM code to check which part is missing. I guess you need to tune timestep carefully.