Made this on complete accident. I think this serves as a practical exercise in what applying a huge amount of lineart style LoRAs to a model until it's almost ruined and then replacing unet blocks IN06, IN07, IN08, OUT05, OUT4, and OUT3 with that of realistic model does.
This model creates a sort of sketchbook like effect to the model. You've got the bones of a semi-realistic model as far as anatomy goes but it looks like someone drew it in with pencil. I call it Perfect Sketch, because basically you can throw any prompt you'd use on Perfect World in here and it'll work.
I can do a more detailed explanation of why this works if anyone's interested but if you're only interested in prompting here are the usage recommendations.
Usage
Ideally you should prompt for realism for best effects. With SketchyAnimeStyle it doesn't really matter, but with SketchStyle it will help a lot. If you don't want to come up with your own prompt, just insert this before your prompt.
(ultra high detail, realistic, best quality, 8k raw, masterpiece)
BREAKAs far as negatives go, keep it simple. Once again, if you don't want to come up with your own prompt, I generally like to use this.
easynegative, (worst quality, low quality:1.4), (monochrome:1.1), watermarkDon't use latent upscaler and if you upscale at 2 or below keep the denoise at 0.4 or below.
With SketchyStyle you should use a upscaler otherwise the facial details will not resolve. With SketchyAnimeStyle it doesn't really matter that much though obviously upscaling is better.
Process
Generally speaking, anything between OU03toOUT5/IN8toIN6 is responsible for segmenting the face and body. I was trying to correct anatomy breakdown after fucking up the weights with loras so I replaced the parts of the unet responsible for compressing and uncompressing the body features with those of the functioning model I started with. This enforces the anatomy of that model into this one to some degree while leaving the fine details intact. I think this could be used as a generalized approach, where you can apply loras to change style and then revert certain blocks to restore anatomy or aesthetic composition. However, this technique is largely experimental and needs more research.
Description
Thanks to rishAI for the comment-question. Helped give me some ideas.
FAQ
Comments (12)
Don't suppose you could do something similar but with a Hentai purpose?
What do you mean by Hentai purpose? I think you could do the same process and get rough lines but if you're looking a certain hentai artist's artstyle you're probably better off utilizing one of the many loras there are there that have been trained already.
So, I'm kinda new to the scene. What's the dif between the Pruned and Safetensor? I see the difference in size, but don't really understand why
Pruned means you killed all the junk data that doesn't affect the output. In this case it's pruned to fp16 which means you get large performance gains out of it since you don't have to have your GPU calculate to fp32. (Be aware that fp16 only works on GPUs from the RTX2000 generation and younger) Safetensor is a file type that can't import and run executable code on your computer (i.e. it's relatively safe to run)
@Jemnite Thanks! I get it now. If you don't mind me asking another noob question, what's the "hash autov2, etc" mean?
@gomezwesno I believe that's because auto1111 used to use a different hashing system. It uses SHA256 now, so that's what hash autov2 should be.
@Jemnite Hashing system? What does that mean?
@gomezwesno It's a general use computing thing.
https://en.wikipedia.org/wiki/Hash_function
Basically you can render any datafile down to a fixed size string or 'hash'. This helps you tell apart files.
@Jemnite Oh! I see now! Thank you very much!
This is my new favorite model, well done
make lora
It seems this model is not doing well with male characters, I wonder if the next version can contain more of this? Thank you！