This is a lora of Suzune from Kisei Juui Suzune (aka. Parasite Doctor Suzune) - Haruki's manga about a woman who bangs all kinds of shit for 120 pages straight.
I love this old-school style, especially her hair and eyes, so I figured I should train a model of her.
Generation settings
1.0-1.2 strength
(Example images were made using the Euler a cfg++ scheduler. You need this extension to replicate them: https://github.com/DenOfEquity/webUI_ExtraSchedulers)
Base appearance
Main tags: KiseiJuuiSuzune, short hair, white hair, large breasts, monochrome, greyscale,
For color: silver hair, brown eyes,
If you want a more hand-drawn look: traditional media, screentones, graphite \(medium\),
Outfits
Standard: black crop top, cleavage cutout, belts, black miniskirt, thigh strap, boots,
With lab coat: lab coat, sleeves rolled up, off-shoulder,
Training details
Base model: Illustrious XL 0.1
84 images, 2 repeats, no flip augmentation, 30 epochs
16 batch size
8 dim, 8 alpha
0.75 unet and text encoder learning rate
Prodigy optimizer, args: "decouple=True weight_decay=0.01 betas=[0.9,0.999] d_coef=1.2 use_bias_correction=True safeguard_warmup=False"
Known issues
Prompting for "faceless male" often gives the guys Suzune's hairstyle. Probably since there are like 0 scenes in the manga that show Suzune and the guy she's banging in the same frame. Add "bald" or something to your prompt to help.
Her thigh strap sometimes show up on both legs, classic stable diffusion
I haven't run into it as a problem, but if you get panels and stuff, put "comic, panels, multiple views, border" in negative, perhaps also "speech bubble, japanese text".
Check out this great model if you want more of the style, but not specifically this character: https://civarchive.com/models/1239311/haruki-manga-style-lycorislora-illustrious
(And upload some images!)
Description
FAQ
Comments (8)
Can you explain how the scheduler works?
I have no clue how the scheduler works from a technical standpoint, but CFG++ schedulers are better at generating smaller details, better at prompt adherence (in my experience) and helps deal with some of the artifacts that Stable Diffusion often suffers from. You can read more about it here: https://cfgpp-diffusion.github.io/
I don't think the CivitAI generator has support for it yet but I'm hoping they implement it soon. I tend to use Euler A CFG++ at the moment.
@Skurkgurkan Thanks. I've read through it and still don't understand how it works though. Are you using the same formula for all works or just this one?
@tkgg2219 I use it for all the images I generate these days, but I only started using it after I uploaded my model of Izayoi Sakuya. Before that, I mainly used Euler A for Illustrious/Pony.
If you're using A1111 or WebUI Forge, all you need to do is go to extensions, click install from url, paste in https://github.com/DenOfEquity/webUI_ExtraSchedulers and then restart A1111. Afterwards, you'll find the CFG++ schedulers when you click the sampling method dropdown. They're made for lower CFG values - I usually use 2.
@Skurkgurkan Not sure what i did wrong but it always messes up the adetailer faces, even with lower CFG
@tkgg2219 Could it be that you've got adetailer set to use a different CFG? I had that issue with hires fix
hey dude, would you be able to do a style lora for this?
there is a style lora already but it's not comparable to the aesthetic and quality this one produces, especially skin and manga shading.
Sure, I can make one. I'll need to rework the dataset so it'll take a bit of time, but that's a good idea :)