Closed zero01101 closed 1 year ago
Well, that's a bit... Sad. What exactly did high res fix did again? Maybe we could emulate old high res fix using two calls to the img2img API?
it'd basically prevent "incorrect" generations with undesirable multiple subjects for images being generated that are larger than the base training data images - previously you'd specify the firstpass width/height which worked great for our needs, but now you specify a scale-by-x value, which would be "downscaled" automatically when openOutpaint got the response due to the reticle/bounding box -_- the liar checkbox tries to sneak that kind of behavior back in by giving SD false values for desired width/height by dividing the dimensions by the hrfix scale factor, so i think it kinda does emulate the old style of operation at least, but there was much grumbling this afternoon when i noticed all my new tricks were instantly worthless lol
Yeah, we need some reworking on the hrfix new features now...
Yeah, we need some reworking on the hrfix new features now...
lol well i think that's what i tried to do with this PR ;)
Okay, will give this a look a bit later tonight.
Oh, nvm, I think I was testing on an older version of the webui.
yeah, that's kinda something i was wondering about; pushing this might make users on an outdated webUI suddenly not have working HRfix, but i can't exactly leave in the previous firstphase
values since the latest webUI console will yell about it every time it's used :/
We can keep this as a parallel branch for now, or we could add a basic preflight on initialization (host connection check) just to check see which parameters cause errors.
lol it seems to be worse than i thought; using previous HRfix implementations with firstphase width/height values passed to SD on newest webUI not only makes it yell firstphase_width/firstphase_height no longer supported; use hr_scale
at you every time, it USES THOSE VALUES as the output resolution 🤦 personally leaning more towards preflight check and informing that webUI needs to be updated if it hasn't been :/
Well, at least it is pretty vocal and easy to detect : )
figured i'd check for use_scale_latent_for_hires_fix
in the startup call to /options since it no longer exists with the new HRfix changes
welp ok seems A1111 has had a fairly extensive HRfix rework that not only completely negates that nifty HRfix lock px slider entirely, but seems to have caused wild havoc in the masses.
this commit completely f&%#$ng reworks the entirety of the previous HRfix system to comply with the new "yeah it's just an upscaler" expectations, does some ugly garbo to the upscalers list routine because we absolutely must have "none" in there as well as completely broken and non-functional "latent" and "latent (nearest)" which aren't even returned in the stupidly non-working-right api call to retrieve upscalers anyway asdl;fkjgas;dkljasdlfgk
yeah i'm a bit frustrated can you tell