This is a lora designed to be put heavily in the negatives to fight the horrible ai slop shiny style of the average r34 and civitai poster
follow my twitter for more https://x.com/AyakonArts
This is a LoRA trained on R34 outputs, generally the place people post the worst over shiny slop you have ever seen, their outputs however are useful as a negative! Simply add the lora at -0.5 to -1 power
Description
doubled dataset size
This is a LoRA trained on R34 outputs, generally the place people post the worst over shiny slop you have ever seen, their outputs however are useful as a negative!
FAQ
Comments (8)
What does this exactly do?
It's trained on imaged with distasteful AI artifacts. The idea is that you use a negative weight on the lora and it will reduce them
At positive weights it replicates those wet/shiny plasticky exaggerated ai generated images that gets posted to the rule 34 sites, at negative weights it does the opposite and does a more flat and grounded style that works pretty well for more realistic styles
legend
This Lora works absolute wonders!
It have a bigger potential. You can make a dataset with only the most generic AI images. Style is not the main problem of AI images. Usually they all look bad because they have no original content - the same poses, scenes, expressions. It would be nice to exclude all these "average" results.
The problem of current version it have negative effect on overall quality. Together with "AI slop" style it learned some extra good things. You can simply process images with various filters to extract only "form" out of images.
This is exactly what i was looking for. I was using a lower CFG scale, but it just makes many details worse for some reason.
And +1, based, there's so much shiny slop and same postures slop with AI, it's crazy
что за лев этот тигр разраб