Tanishi (hentai model)

A model with Tanishi artstyle in mind. There is already a Lora if you want that (here). The Lora is not made by me. Trained on AnyHentai as base model.

What does Tanishi style add:

-Pubic hair added easily (anus,vagina) in images
-Tanish arstyle armpit hair added automatically if there is armpit in the image

-Easier to obtain Thick,plump, big breasts, saggy breasts images

-Clothes are tight even if I don’t prompt that

-Super black skin(12.5% chance to get one image like that)

-Easier to obtain black and white images compared to other hentai models

Negative prompt: (speech bubble:1.3),(text:1.3),(logo:1.3),(signature:1.3)

Important to read for Negative prompt
You might want to add in certain images/seeds negative prompts such as: (pussy juice) or (naked) or use weight like (naked:1.3). I use Low cfg scale so I say this because It tend to generate so NSFW in certain cases, that you actually don’t want that, so for low cfg scale numbers, add strong negative prompt to avoid that.

Full Negative prompt: easynegative, badhandv4, lr, beard, curly, veiny, muscular, ((bodybuilder)), facial hair, mustache, lowres, fewer fingers, out of frame, jpeg artifacts, ugly, duplicate, morbid,mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers,(speech bubble:1.3),(text:1.3),(logo:1.3),(signature:1.3)

Settings
It should work with your usual settings but here are mine:

• Sampler: Euler a (is what I use but works with others as well from what I saw)

• Size: 512-512 / 512-744 (didn’t tried with 512-768, but I assume it works)

• Clip skip: 2 (seems clip skip1 images are ok? I guess, but I main “2”)

• Sampling Steps: 20-30 (I use 30 all the time)

• Hiresfix upscaler: R-ESRGAN 4x Anime6B on 2x, steps: 20, denoise 0.5 (I didn’t test higher or lower than 0.5).

• Cfg scale: 6-7 (didn’t tried higher or lower)

Why make a model if there is a Lora already? (#rant)
– Because I find Lora to be very stiff, yes Lora are great, and I use them all the time. But I wanted to make “black skin” or “black and white” images. Can the lora do that? yeah… It kinda struggle to do that. And I think it’s because it mainly had only white girls in training. If there is a booru tag that get that specific thing I want, I am too lazy to find that specific tag, to activate Lora’s potential. (‘why don’t use multiple Lora to get what you want’, what if I want 100 unique things, that require 100 unique Lora, if a model can do those 100 things, i’d rather spend the time to make a model than spending the time to get all the Lora I want)

You said earlier it’s important to read that text for negative prompt but wasn’t important tho?
– Yeah I understand that 90% of people here already know to use models, I wrote that in case people copy me and use my cfg scale and they get half naked model even if they wrote: jeans,shirt,fully clothed. I know that higher cfg scale should in theory fix this issue, But I usually enjoy when I let the model be the artist and surprise me, so I still want that randomness and not complying to my full prompt, but at the same time avoid certain elements, so the weight in negative prompt should fix that

How Many Images I used for training?
– 166, And I edited around 10. I don’t use mask to get a character out of an image, I like to get the background as well, I feel like if the character get pulled only, then the background will get added by whatever model I train mine on, and I don’t want that type background (janky girl with realistic background for e.g). Plus there is always merging if people want same artstyle and improve stuff such as Background or body,etc.