If you enjoyed this model, why don't you check out my other works? I put a lot of effort into correctly recreating artist styles and making the occasional character LoRA.😊
This model was trained using the collab/guide by holostrawberry [Guide] Make your own Loras, easy and free - colabs | Stable Diffusion Other | Civitai
Model used in examples:
model.ckpt · a1079602570/animefull-final-pruned at main (huggingface.co)
I am once again open for suggestions for other styles, SFW and NSFW.
You can always drop model suggestions in the comments, both for styles that are missing on civit or those that are outdated.
About the artist:
We are back to hentai. In this case the dataset was around 50/50 colorized and monochrome.
Twitter: (1) ヘリを📚C102日曜東C13b (@herioscope) / X (twitter.com)
Herio is a well known artist with a preference for large breasts and incredibly expressive faces and poses. Their content is generally really wholesome and high effort.
(That is something that can't be translated into this model)
Weights & Prompts
The only necessary prompt is:
HerioUnsurprisingly this model works really well with nsfw and explicit content. Sex positions in particular tend to work better than most base models. This does not extend to detailed genitalia. Censorship is to blame in that department.
Use a weight of 0.8-1.2, adjust to taste and depending on your checkpoint.
As usual I have included an XYZ plot for reference.
The model can consistently create colorized and monochrome images depending on your prompts. Adding any color specifications usually guarantees full color, "monochrome" or "greyscale" moves it closer to the manga style.
Please leave a review if you enjoyed this model or have problems with it. This helps me with improving future models and finetuning them in general. Give it a try and throw some images down below! 👇
Description
FAQ
Comments (4)
I'm gonna take a look but it's probably gonna take a while.
I can't make nsfw very well.
Can you help me?
As a baseline if you don't prefer any specific checkpoint I recommend you use the same I utilized for the examples. I've added the huggingface link in the description (it's a lot bigger than most others as it is pretty much the og anime checkpoint,, the pruned version by the same creator might work as well)
For nsfw it heavily depends on what you want to do. Straight up nudity with different poses shouldn't be any problem. Using "nude" or "completely nude" can be helped along by giving more details like "nipples", "navel". It is rare to not get a good result in 4 images.
Actual intercourse or many specific other concepts are a lot more challenging. I can't really train the model to include poses that aren't in the source material multiple times and if the base model doesn't really know how to depict a concept, you need to use additional LoRA.
Things that always tend to work (but you usually need more images to get something that looks right - anatomy is hard for ai) are typical positions, "cowgirl", "doggystyle", "missionary" - add stuff like "sex, hetero, vaginal/anal" and "1boy, 1girl". "pov" also works really consistently. I also recommend increasing their weights to 1.3 or more. Stay with few prompts and only specify additional things if they should be there but are missing. I have seen too much futanari because I specified "penis" in addition to the standard nsfw tags.
Last thing to be aware of: aspect ratio has a massive effect.
My general workflow:
Basic prompt -> 4 images -> check if it works at all ( -> if not switch around aspect ratio) -> add details, and backgrounds -> 4 images (or more, depends on how strong your GPU is) -> choose best layout and repeat that seed with different sample steps (12,16,20,24,30 in my case. Somewhere around 30 changes are minor but can be significant - hand in particular) -> choose best of those and repeat that with high-res fix (half the sampling steps, 0.5 denoising strenght) -> if that ruins any aspect go lower with both.
Something I can't do and might be worth it: Mage a ton of images always running highres-fix at 0.7 denoising strength and the largest upscale your ram can support. At that point you can't work from non-hi-res to hi-res because it is that different, but the detail is comparatively that much better.
I might have to rewrite all of the above and make a post I can link to 😂