I used a dataset of around ~315 images for v2 and I used Source checkpoint: AnyHentai as the base model (the 2.0 one, was +7GB). Lora here.
あなたがその著者であり、これを読んでいるなら、このモデルを使ってコンテンツを今までよりも速く作成し、より多くのコンテンツを公開してください。不正確な部分を描き直すか、それらをアイデアを得るために使ってください。
PS: This is just a NSFW model made after Akatsuki Souken circle to output images with similar style and body type that the mangas have. I made this model to share author's name to more people that might be interested in it, and hoping the author finds the model in the future and uses it to get ideas, or use it in his own work to produce more content.
Settings:
Sampler: Euler a / DMP++ SDE Karras it's what I use
Size: 512-512 (didn't test for other resolutions)
Clip skip: 2 (I used clip skip 2 mostly for image generation, but initially I was using clip skip 1 and seems fine)
Sampling Steps: 20-30 (I use mainly 30, but the coherency drops down at 20 for certain complex prompts, test for yourself, sometimes I go 28-35 if I find a seed I like the output)
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 (6-11 I use)
Prompts about the x-ray/manga panels:
(manga panels) - produce cutouts in the image and usually if there is genitalia in the image will zoom on it
(x-ray) / (xray) - produce cutouts in the image and if there is intercourse will show the classic x-ray of akatsuki souken
(manga panels), (x-ray) - produce (sometimes) a mix of both
sex - produce blowjob mainly as a standalone word (by adding ,mating, thrusting, couple, you get sex, tends to produce sexual NSFW content)
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,
I used that one for all images but you can change/edit, and make it better. Easynegative and badhandv4 are textual inversions (embeddings).
How I create prompts:
- I go in chatgpt, I give it one or three examples of prompts that I already have and I tell it:
"I am generating images for an Image generator text, the text it needs is special and the template for it is: (word,word,word),(word,word,word),...,(word,word,word), (details:1.3),(masterpiece:1.3).
- The reason I use (word,word,word) is so I box things together and be able to change easier, (green eyes,makeup,glasses),(pendulous breasts,erect nipple,pasties).
- Then chatgpt just generate for me easy prompts, I use same negative prompt
What's new in v2:
- The training, I carefully edited in photoshop around +200 images (probably 250). I did edit and made sure that there are no smearing like the images in V1. Most images should end up more crisp. The amount of smearing should be 3-5% max, so not noticeable my opinion (compared to what it used to be). Total img used around ~315(60-70 for v1 as comparison).
- The nipples look better. In v1 a lot of nipples had double nipple or it was so erect it launched itself out of the body.
- The color and style changes. I used around half of his works that he ever published (I think, if not 50% then 25% for sure). Because there are volumes from years ago and more recent ones, the "style" he used to draw changed so this will be seen in the final result. It won't feel probably like Akatsuki souken the final output, but more like a "high quality" one. Which might be good or bad, depends on ur taste. I didn't try combining v1 and v2 (maybe that could make the style more like v1 if that's ur taste).
- Downgrades: 1)
Because the volume of images is higher, the amount of "internal shots" that I put are so many, and look different so it's not as crispy as v1. V1 had only a few images, so now I'd say when a good result occur, it's usually better than v1, but v2 is inaccurate, so most x-ray might be bad.
2) Also the "sex" prompt produces a blowjob scene instead of a sex scene. I think it's because there were too many blowjob scenes I added in there. So the conclusion is it's hard to obtain NSFW stuff. I would probably use "lora" or control net to obtain certain things that used to be easier in v1. I think combining v1 with v2 or v2 with some other heavy hentai model might solve this issue.
My advice if you want something weird or whatever are your needs is combine models.
1) Comparison Image for v2 combined with a custom model I have.
2) Comparison Image for v2 combined with same custom model I have
For 2nd I think I used "huge pendulous saggy breasts" or something and it got that crazy result XD. Either way I think the =.= way to go is combining models.
Description
This is my first model I trained and I used a dataset of around +60 images and I used Source checkpoint: AnyHentai as the base model (the 2.0 one, was +7GB).
あなたがその著者であり、これを読んでいるなら、このモデルを使ってコンテンツを今までよりも速く作成し、より多くのコンテンツを公開してください。不正確な部分を描き直すか、それらをアイデアを得るために使ってください。
TL;DR It's a hentai model trained on images from a Circle called Akatsuki Souken , it should output images with that style in mind
Q: Why did you do it?
A: I did this model so others can enjoy more content and find out more about the author. But personally I hope the author finds the model and use it himself to produce more content, he would only need to redraw certain parts, and make way more content than before
Settings:
Sampler: Euler a / DMP++ SDE Karras it's what I used, but you can test yourself (others samplers give interesting results as well)
Size: 512-512 works the best, ocasionally I tested with 512-768 and I get mix results.
Clip skip: 2 (It seems to improve everything, and reduce the cutouts panels that appear in images and improve eye/face/easier to get a background)
Sampling Steps: 20-30 (I also tested with 150, but my first 60% tests I used 20 without realizing then I switched to 30 mainly)
Hiresfix upscaler: 4x-UltraSharp on 2x, steps: 20, denoise 0.5 (0.2-0.5) seems ok results (took it from Olivio Sarikas) . Latent and R-ESRGAN 4x Anime6B also similar results
Cfg scale: 6 (6-11 I used. I noticed the higher I go the worse results I get)
Prompts about the x-ray/manga panels:
(manga panels) - produce cutouts in the image and usually if there is genitalia in the image will zoom on it
(x-ray) / (xray) - produce cutouts in the image and if there is intercourse will show the classic x-ray of akatsuki souken, kinda distorted since the dataset was small
(manga panels), (x-ray) - produce (sometimes) a mix of both
sex - produce intercourse (usually no cutouts)
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,
I used that one for all images but you can change/edit, and make it better. Easynegative and badhandv4 are textual inversions (embeddings).
DataSet
I trained with 4 concepts the prompts I used(if I remember correctly) for those are
- akatsuki souken plump girl (around 20 img) [images with girls]
- akatsuki souken plump girl sex (around 21 img) [images with girls that make sex or blowjob and no manga panel in the image]
- akatsuki souken plump girl x-ray (around 16 img) [all images contain a girl that make sex and have internal view with the womb as manga panels]
- akatsuki souken plump girl manga panels (around 6 img) [the images contain image with the girl and 1-2 manga panel in the image that showcase either her boobs or anus or zoomed in on vag
Images training dataset
The images I used to train are RAW textless versions that I did edit out bits and cutouts in photoshop. I am a noob at editing photos so there are a lot of artifacts in the original images I did train. Sometimes u might see it yourself if certain "cutouts" or "backgrounds" look funky in the generations. I originally planned to add way more than 70 images (+1000), but it take so long to edit images that I don't think I will do that too soon. All images I resized them to 512x512 . I trained a model without modifying the manga panels but you would get weird results such as "male heads" and manga cutouts in scenes that you usually don't want which makes results unusable. So editing in photoshop is a necessity for most images (that's why only 70 img)
Machine used to train
I used runpod stable difussion 5:0 template and A6000 to train it and test (I have 2019 macOS can't train on my laptop sadly)
How easy is to use it: 5/10. Personally I am a noob and takes me a while to get the image I have in mind, Overall I was able to get almost anything I wanted from him that I like tweaking settings.
TIP:
I actually noticed that I have a hard time getting "crisp" results by myself, to fix that I went to chatGPT (3.5) and I asked him for 10 ideas such as:
"A surfer chick wearing a colorful wetsuit, riding the waves at a beautiful beach in Hawaii."
Then I learned chatGPT to spit out a version like this:
"The surfer chick rides the waves at a beautiful beach in Hawaii, dressed in a (vibrant) wetsuit that catches the eye with its (colorful) design. The (form-fitting) suit hugs her (athletic) figure, allowing for freedom of movement as she gracefully navigates the waves. Her (long) hair is tied back in a (messy) bun, with a (sun-kissed) complexion that adds to her (beachy) vibe. The (cool) water sprays her face as she rides the waves with (skill) and (determination), navigating the (challenging) surf with ease. The (sleek) surfboard is an extension of her body, as she deftly maneuvers it with (precision) and (grace). The (scenic) beauty of the beach and the (salty) air adds to the (thrill) of the ride, making for an (unforgettable) experience. Despite the (dangerous) nature of the sport, the surfer chick exudes (fearlessness) and (adventure), fully embracing the thrill of the ride."
I called writing in that style "akatsuki style", then I would give any sentence to chatGPT and tell it to write in akatsuki style and would output a version of text like that. By doing like this I could get quite good results (non nsfw). Then all u have to do is copy that text into the a1111 as prompt (or in chatGPT and get yourself ur own style, works with other models as well this chatGPT idea)
PS: Try adding --no-half-vae in command args, runpod dreambooth is tricky to setup, and my model sometimes fail to make an image, I noticed "--no-half-vae" reduced the amount of failures significantly, with this only 1 image out of 20 fails and give an error, there is a way to disable the checkup (which should stop the errors all together, but I didn't managed yet to do that)
FAQ
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.