AIroticArt™ Notes

This is where we share some of our notes, models, and tips with the AI art community.

AIroticArt's Vulva Model v1.0

It's needless to say that standard AI models usually can't generate female genital anatomy very well. They were just not trained on that data. At best, the models might produce the vulval cleft, where the labia majora meet, also called the pudendal cleft, but that is about all. There is so much more to the vulva, this beautiful female organ, which is actually several organs. There are the labia majora and labia minora, the clitoral hood that usually covers the clitoris (or clitoral glans), the urethra, the vagina (and vaginal orifice), the mons pubis, perineum, asshole/anus, etc. But all of these parts are usually hidden or simply absent in present AI models.

We wanted to change that, so we trained a model with Dreambooth that specializes in detailed closeups of anatomically correct female genitalia, i.e. all the parts of the vulva, including hairy and hairless (shaved), wet/moist and dry, and all different shapes, sizes, and colors of the vulva. And we're giving it back to the community.

Examples

Here are some examples of what v1.0 model can do. These are direct from txt2img, no inpainting or img2img, and only minor cherry-picking.

Download normal model

2GB download:
https://civitai.com/models/2389/airoticarts-vulva-model

Mega mirror:
https://mega.nz/file/JX9wjAhC#P70jkCh77CD_m_LxwqCViqlY5FkdfJ39tF4lXX0-rPY

model hash: 0c2b65c4
md5: 717393a25bc36e0ee799c19c3dbbacf3

Download inpainting model

This model may do much better when used as an inpainting model, to correct vulva anatomy on other images in inpainting mode. This inpainting model was generated by doing a simple checkpoint merger, with "add difference" interpolation, sd1.5-inpainting + (vulva1.0 - sd1.5-emaonly) * 1.

2GB download:
https://civitai.com/models/2389/airoticarts-vulva-model
(be sure to select the 1.0 inpainting model version)

Mega mirror:
https://mega.nz/file/pHUxhAoT#q2OVpmEbLBc8MaGy-i54IRWJr6KzjUSw72nOSKm5aJU

model hash: 045fbd65
md5: 0abfc36312c5a7a98ce81699a2d5ee1d

Training Dataset and Captions

Here is what the images and captions used in training were like:

• 60 images, with detailed captions like "a photo of a nude naked woman's vulvdet, vulva, pussy, shaved, hairless, labia majora, labia minora, clitoral hood, moist, wet, goosebumps, ass, anus, asshole"

Training Parameters

Dreambooth training (Shivam)
60 images, manually edited for clarity
512x512 resolution
BLIP initialized captions, then manually edited to add detail
DDIM scheduler
SD 1.5 pruned-emaonly base model
fp16 precision
No class/regularization images (no prior preservation)
4,000 steps (training text encoder 100%)
LR 2e-06

Example Prompts

You can simply use "a nude naked woman" or "a nude woman's vulva" in your prompt and get some good results.

You may add more details as noted in the caption above, including vulva, pussy, hairy, hairless, shaved, wet, moist, labia majora, labia minora, vagina, clitoral hood, clitoris, ass, asshole, anus, legs spread, butt, indoors, outdoors, skin tone, open, closed, gaping, etc. You may also add the unique instance keyword vulvdet (vulva detail) to more directly specify this training dataset.

Limitations

For this model we trained closeup shots mostly, so the model might not be able to produce more wide angle shots of women. Sometimes it still doesn't get the genitalia correct, and may duplicate labia or anus. At times the clitoral hood can look too sharp or rigid or too large. The clitoris often is still hiding. A greater variety of training images, with a variety of different anatomy, settings, poses, and crops, would produce more variety in inferencing. Using class images with prior preservation might allow the model to produce wider angle images. This model was trained mostly on closeups, so might be best used for inpainting fixes. Training could be done on more ethnicities. Full precision could be used, as well as the full ema+non-ema base model. These are areas for improvement in future versions of the model.

Enjoy!

AIroticArt's Penis Model v1.0

The present state of standard AI models is that they do not produce male genitalia well, if at all. Generating nude males usually results in eunuchs, or even completely emasculated figures. The penis and/or testicles are missing or completely disfigured in sometimes gruesome ways. Ironically, by trying to avoid training models on accurate human anatomy, they have produced models that actively disfigure humans in disturbing ways. Michelangelo is rolling in his grave.

We wanted to change that, so we trained a model with Dreambooth that specializes in detailed closeups of anatomically correct male genitalia, i.e. penises, testicles/scrotums, pubic hair or shaved, both flaccid and erect. And we're giving it back to the community.

Examples

Here are some examples of what v1.0 model can do. These are direct from txt2img, no inpainting or img2img, and only minor cherry-picking.

Download normal model

2GB download:
https://civitai.com/models/1245/airoticarts-penis-model

Anon mirror:
https://anonfiles.com/n2m9j9I8yf/airoticart_penflac-penerec-1_19100_ckpt

Mega mirror:
https://mega.nz/file/JOUy0ABJ#R2B537N-rM3hTvv47ZAi6JntS3ShQEkSanv4PIyIgtU

model hash: e5a80698
md5: c94066c5dd9335f8691e09ed87113cb9

Download inpainting model

This model may do much better when used as an inpainting model, to correct penis anatomy on other images in inpainting mode. This inpainting model was generated by doing a simple checkpoint merger, with "add difference" interpolation, sd1.5-inpainting + (penis1.0 - sd1.5-emaonly) * 1.

2GB download:
https://civitai.com/models/1245/airoticarts-penis-model
(be sure to select the 1.0 inpainting model)

Anon mirror:
https://anonfiles.com/Y6Bb3aLby8/airoticart_penflac-penerec-1_19100-inpainting_ckpt

Mega mirror:
https://mega.nz/file/FasiRRzR#wXzOpLz0zW3GnPJh-Hp67v7m6EIwF7ZcxUtY08erTRo

model hash: 2dafa8e1
md5: f3091ffb5f80a848d73df1832ce072e9

Training Dataset and Captions

There were two basic kinds of images and captions used in training, as two separate concepts:

• 30 flaccid images, with detailed captions like "a medium full shot of a nude naked man with a hairy body standing in a room with a lamp and a picture frame on the wall, flaccid penis, frontal, folding arms, pubic hair, circumcised, glans penis, scrotum with testicles, head and face out of frame, penflac"
• 30 erect images, with detailed captions like "a closeup of a nude naked man's erect penis, bare legs, no shirt, outdoors, shaft, cock, dick, from the front, scrotum, testicles, shaved, no hair, smooth skin, glans penis, frenulum, penerec"

Training Parameters

Dreambooth training (Shivam)
60 images (30 flaccid, 30 erect), manually edited for clarity
512x512 resolution
BLIP initialized captions, then manually edited to add detail
DDIM scheduler
SD 1.5 pruned-emaonly base model
fp16 precision
600 class/regularization images, 512x512 (generated with "a photo of a nude naked man")
19,100 steps (training text encoder only at start, ~1%)
LR 2e-06

Example Prompts

You can simply use "a nude naked man" or "a nude man's penis" in your prompt and get some good results.

You may add more details as noted in the captions above, including flaccid or erect, arm positions, the environment, indoors or outdoors, size of shot (full shot, medium full shot, closeup), etc. You may also add the unique instance keyword penflac or penerec to more directly specify a flaccid or erect penis, respectively.

Limitations

For this model we trained circumcised penises, and in the future we want to include uncircumcised penises with foreskin intact as well (erect penises look similar already). The model can still generate mangled or duplicated genitalia that can be disturbing. Sometimes indoor shots produce too much film grain from low-quality training images. Sometimes prompting for flaccid will generate erect, and vice versa. A greater variety of training images, with a variety of different anatomy, settings, poses, and crops, would produce more variety in inferencing. This model was trained mostly on closeups, so might be best used for inpainting fixes. Training could be done on more ethnicities. Full precision could be used, as well as the full ema+non-ema base model. These are areas for improvement in future versions of the model.

Enjoy!

AIroticArt's Sidelyer Model v1.0

Most present AI models cannot generate images of humans lying down very well, especially on their side. People are almost always standing, or maybe sitting. If you attempt to generate people lying down or "sidelyers" you often get mangled bodies that are hardly recognized as human. The models just haven't been trained well on these body positions.

We wanted to change that, so we trained a model with Dreambooth that specializes in sidelying, front and back, the classic nude pose. And we're giving it back to the community.

Examples

Here are some examples of what v1.0 model can do. These are direct from txt2img, no inpainting or img2img, and only minor cherry-picking.

Download

2GB download:
https://civitai.com/models/1243/airoticarts-sidelyer-model

Anon mirror:
https://anonfiles.com/Z7iaBdH9y5/airoticart_sidelyer1_28545_ckpt

Mega mirror:
https://mega.nz/file/4XthXTBY#hCEJgD0PmV4LWP3fOOlisksBEEsr_o0yDsDM5fvveiQ

model hash: fd907af7
md5: fbd0d2a711685bf8b9dd06a4b8472976

Training Dataset and Captions

There were two basic kinds of images and captions used in training, as two separate concepts:

• 141 frontal images, with detailed captions like "a full shot of a naked nude brunette woman with long hair lying on her right side on a white bed, leaning on her right arm, her left arm crossed in front, both legs bent at the knee, left leg and knee vertical, pubic hair, looking at the camera, white background, sidelyerfront"
• 128 back images, with detailed captions like "a medium full shot of a nude naked blond woman lying on her right side on a wooden dock by a lake with a reflection of trees in the water behind her, legs straight, left leg in front of her, back side, butt, ass, pussy, wavy hair, head raised, facing away from camera, sidelyerback, soft diffuse lighting"

Training Parameters

Dreambooth training (Shivam)
269 images (141 front, 128 back), manually edited for clarity
512x512 resolution
BLIP initialized captions, then manually edited to add detail
DDIM scheduler
f222 base model
fp16 precision
1,000 class/regularization images, 512x512 (generated with "a color photo of a nude naked woman")
28,545 steps (training text encoder only at start and finish, ~5%)
LR 5e-06

Example Prompts

You can simply use "a nude woman lying on her side" in your prompt and get some good results.

You may add more details as noted in the captions above, including front or back side, arm positions, hair color, where they are looking (at the camera, away from the camera), the environment, indoors or outdoors, size of shot (full shot, medium full shot, closeup), etc. You may also add the unique instance keyword sidelyerfront or sidelyerback to more directly specify a front or back view, respectively.

Limitations

The model isn't perfect, as no model is. Feet are still often mangled. Tattoos in the dataset look like burn marks. Sometimes there is repetition in the upper and lower regions of the image, as some of the dataset was expanded vertically to fit 512x512 resolution. Sometimes the bodies are still mangled. This training was only on females, no males. Training could be done on more ethnicities. Full precision could be used. A slower learning rate would likely produce a higher quality model. These are some of the areas for improvement in future versions of the model.

Enjoy!