How to get results that can change to other/better results?
I added some code to write intermediate steps and I notice that when I increase the number of steps, I still get a lot of noise until the last few steps.
In other demos I saw stable diffusion converging to one good looking solution in lets say 30 steps and then changing the solution to other good looking solutions (e.g. by adding and removing details) in the next 20 steps.
Does somebody know how to achieve a result like this?
Increasing the number of steps only means that the denoising process occurs more gradually, leading to stabler (and often higher quality) results, to a point (eventually increasing # steps doesn't do anything).
The demos you're referring to are probably using the img2img script, where you initialize the denoising with a particular image and guide it to a different output with some prompt