Open IDKiro opened 1 year ago
I conducted experiments applying DDIM and DPM-Solver and found that they performed very poorly. After debugging, I found that ODE-based samplers inherently couldn't reconstruct images similar to the conditional images. I tried setting the eta of DDIM to 1.0, and obtained results similar to those from the Spaced Sampler. Did you also remove DDIM based on similar experimental observations?
Yes! We find that the DDIM sampler will produce unpleasant results and we don't know what the reason is.
This is an interesting question. If you come across any new findings in the future, I hope you'll let me know. I'm also planning on exploring this area further in the upcoming period.
I noticed that you've removed the DDIM option in the current version of the code, even though it didn't seem to work in the initial version. Sampling efficiency is one of the obstacles to practicality, especially for high-resolution images. Will DDIM and DPM-Solver samplers be supported in the future?