Closed mspronesti closed 1 year ago
The failed check has to do with a file I haven't worked on in this PR. I think it's due to the fact that the latest version of clippy requires inlined variables in format strings. Should be easy to solve with cargo clippy --fix
.
Thanks for the fix, and no worries re the clippy lint, I'll just run a clippy fix everywhere.
When I implemented the DDPM Scheduler (#21), I noticed that nothing below 1000 inference steps produced a "meaningful" image (with 1000 training timesteps). Basically I only had noise back. However, since this behavior was common to HF python version, we agreed it was "normal". Nevertheless, this morning HF people pubblished a fix to this scheduler, which contained in fact a bug when used with a number of inference steps smaller the number of training steps. This PR aims at porting this fix to diffusers-rs.
Example of produced images with 30 steps: