This is a New Style Lora I have created, it is suitable for cartoon especially manga styles.
Interesting thing about this model is it's sort of balance between realistic and cartoon, therefore I suggest you to go and play with it see which one you like. The training data is mostly from the manga images I have found online :)
I will suggest using 0.7 to 0.9 weight for this Lora. Using weight above 1 can look over fitting.
In the example images I used the same prompt and experiment with different Models.
This Lora is more stylized therefore the hand part may need additional fixes or extra care.
Model: Most known models
Prompts:
Male images:
monochrome), (manga style), (manga line shades), a man with a large body and a large head, abs, with in frame, ultra-detailed, 8k, masterpiece, detailed face, (volumetric lighting), detailed eyes, (ambient light), realistic shadows, (fullbody:1), (cowboy shot), jean pants, looking at viewer, body well:1, leg well:1, arm well:1 , (GoodhandsV3)
Neg: (easynegative), (hands poor), (fingers poor:1), (badquality), (bad anatomy), (nsfw), (legs poor), (inaccurate limb:1), (2girls), bad composition, inaccurate eyes, (extra hands:1.2), (inaccurate limb:1.2), (extra arms:1.2), (bad quality), bad face, (bad eyes), deformed, ((deformed fingers:1.2)), ((extra fingers, extra limbs)), bad hands, bad fingers
Steps: 31, Sampler: Euler a, CFG scale: 9, Size: 512x768, Denoising strength: 0.37, Clip skip: 2
Female images:
(monochrome), manga style, (hand well), (finger well), (long braided hair), (bangs), with in frame, ultra-detailed, 8k, masterpiece, beautiful detailed face, (volumetric lighting), beautiful detailed eyes, (ambient light), realistic shadows, (fullbody:1), (1girl), (cowboy shot), shirts, pants, looking at viewer, perfect body, body well:1, leg well:1, arm well:1 , (GoodhandsV3)
Neg: (easynegative), (hands poor), (fingers poor:1), (badquality), (bad anatomy), (nsfw), (legs poor), (inaccurate limb:1), (2girls), bad composition, inaccurate eyes, (extra hands:1.2), (inaccurate limb:1.2), (extra arms:1.2), (bad quality), bad face, (bad eyes), deformed, ((deformed fingers:1.2)), ((extra fingers, extra limbs)), bad hands, bad fingers
Steps: 31, Sampler: Euler a, CFG scale: 9, Size: 512x768, Denoising strength: 0.37, Clip skip: 2
This is the version I believe to have the best result right now, but I do have other test models I think are good as well. I will leave them in my Patreon welcome to go there and try it out :)
https://www.patreon.com/posts/boichi-style-82229695
Description
I have retrained and updated the dataset for this Lora, increasing the quality and quantity on images.
The improvements you may see in this version, the shadow will have more detailed fine lines, the anatomy will be more accurate, more manga stylized elements included.
FAQ
Comments (12)
if you can do the same for tokyo ghoul manga, it's much trickier though
Hmm, sure I can give it a try with Tokyo Ghoul manga style, the style does look quite messy and abstract for sure. I could be fun see how SD handle this hahaha.
@EnthusiastAI yeah most people consider it to be one of the top manga styles, but as you said it can be quite abstract, a lot of dark contrasting shapes and hatching patterns lol. That would be cool, I as well just wanna see if manga training is substantial since most people say it's not, but I still think it is
@cyan8 Hey mate your long awaited tokyo ghoul manga style lora is here, have fun :) The training is possible but you just have to do a lot of pre-process to the image before you can use them.
https://civitai.com/models/92028/ghoul-style
@EnthusiastAI hey, yeah I saw it and left a review already :)
I'll be trying it out
Actually I'll be hooking you up with a dataset I'm working on, just need a few more days
I can't reproduce Hirose Mai from the manga origin well
Can't you put out the characteristic long twin tails?
Hmm this must be because the majority of my data set is from the manga Sun Ken Rock drawn by Boichi. Even tho the character you are mentioning is drawn from the same artist, the data set does not include this. The solution to fix this issues will be adding another Lora or textual inversion train with the character Hirose Mai or Long twin tail hairs. Or you can try to lower the weight in this Lora and emphasize on the word (long twin tails) in the Prompts.
So in my new image post, it is recommended to lower the weight strength to around 0.35 to 0.5 if you want to add specific features to this Lora model. This will give SD more freedom to generate based on your prompts. However, reducing the weight setting may result in the loss of some details from the original style, so feel free to experiment with different settings to find what suits you best.
This is the prompt I used:
((((long twin tail hair:1)))), ((long twin tails)), Hirose Mai, (monochrome), manga style, (hand well), (finger well), (rabbit hair clip), with in frame, ultra-detailed, 8k, masterpiece, beautiful detailed face, no names, (volumetric lighting), beautiful detailed eyes, (ambient light), realistic shadows, (fullbody:1), (1girl), (cowboy shot), shirts, pants, looking at viewer, perfect body, body well:1, leg well:1, arm well:1 , (GoodhandsV3)
thanks i will try
Very cool! Would really like to see even more manga styles LORA, maybe Nabaka (Seven Deadly Sins) and Kishiro (Battle angel alita, the original, not the movie).
Hmm those are interesting manga styles, I could take a look into those.
Interesting model!
I think it would be helpful if you were to note which parts of your promt are textual inversions [make them into a different color or in Bold?] and also a link would be useful (I wasn't able to find GoodhandsV3, for example).
Thanks :)