Add option to use noisy latent for hires fix

[brkirch]

Description

This PR adds an option, Hires noisy latent, which sends a noisy latent for the highres pass (without renoising it), which is obtained by producing sigmas for the number of steps for the first pass plus the number of steps for the second pass and then trimming the excess sigmas. Currently this is only implemented for k-diffusion samplers, and it doesn't appear to work correctly for all of them. Also this doesn't really work for increasing resolution, the intended usage for now is sending from the SD XL base model to the refiner without changing resolution.

This should allow for more closely matching the official SD XL workflow when used with #12181.

Screenshots/videos:

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[AUTOMATIC1111]

oh no
I confused this with one of my branches
I didn't meant to delete this

[AUTOMATIC1111]

No, I think that was my branch...

Just to clarify, I didn't willingly close this, it was done automatically by github when i deleted the branch I merged.

[AUTOMATIC1111]

Looking at the code, I think the correct way to do this is not to add noise after you've finished sampling, but to stop sampling early instead.

My vision for refiner is to make this possible separately from hires fix.

[brkirch]

@AUTOMATIC1111 Yes, I can confirm that I made the process needlessly complicated. It looks like a working approach is to start sampling for (first pass steps + second pass steps), stop after first pass steps, then run img2img with the refiner for second pass steps without adding noise and also setting denoising strength to (second pass steps) / (first pass steps + second pass steps).