blepping
commented
9 hours ago sorry, i don't have an example workflow published right now because i'm still frequently making changes to the project. it is relatively straightforward to use though.
if you don't connect sigmas or samplers, it'll do 15 steps of Karras at 0.3 denoise with Euler. if you don't connect a reference image, it will generate one for you based on the latent you connect to the sampler. for example:
this will generate an initial 512x512 reference image (via 20 Karras steps), then do two 2x upscale passes afterward - so 512x512 to 1024x1024 then 1024x1024 to 2048x2048.
as for what to try other than the defaults, i have some suggestions in the main README. most of my testing is with SD15, i like something like this:
you can also use iteration overrides (see the README) to do stuff like switch schedules/parameters. i've found using an upscale model seems to make the most difference going from low res to medium and it's less of a benefit when you're already at a decent resolution: it's possible to use the upscale model just for the first iteration (possibly also image guidance) and then switch to simple scaling/latent guidance for subsequent iterations. you also can do more iterations with a lower scale factor.
Hello found your repo, looks nice , is there any example workflow?