>>63613357That's the problem, it doesn't do anything for the composition but for some reason they come out nice on the first pass.
The workflow itself just has a weird uspcaling method where it processes a 512x512 image with controlnet lineart, then upscales the image, downscales it and processes it again with the controlnet as conditioning and the upscaled-downscaled image as input.
The result is awful and I need to work on it but there is something wrong with this negative conditioning because it turns out the best results happen when I leave it with "(revealing clothing, cleavage, sexy:1.5), (exposed skin:1.5)" in the negatives but these images don't seem to negate any of that.
Here's the workflow. SD1.5
https://files.catbox.moe/gzmgi1.pngIt must be something about that negative combo because if I remove the non-embedding parts of it, the image goes bad.
The style loras might be irrelevant. I've seen it pump out good results even without them. They're just a combo I got from the Mako Sameshima anon.
