search

devilismyfriend / StableTuner

Finetuning SD in style.
GNU Affero General Public License v3.0
666 stars 51 forks source link

Change how noise is added to latents to better learn contrasts/strong colors #101

Closed allenbenz closed 1 year ago

allenbenz commented 1 year ago

https://github.com/devilismyfriend/StableTuner/issues/100

https://www.crosslabs.org/blog/diffusion-with-offset-noise

Ran across this article, tried it, got better "low frequency" color reproduction when doing inference. Basically got obviously better looking chiaroscuro and colors seemed more accurate. Has a marginal impact on performance, didn't notice any coherency issues.

Still, maybe this should be behind a toggle?

Figured someone else should try this to make sure I'm not imagining the improvement.

ProGamerGov commented 1 year ago

It should probably be in an if else statement connected to a UI toggle, with the original behavior as default.

ProGamerGov commented 1 year ago

Looks like the UI option isn't showing up in the UI.