iFAP - Softcore

Update for Version 2 - see file version info for details. All sample images created without the use of adetailer to show the good and the bad. Adetailer helps things quite a bit.

Version 2 features dramatically improved prompt ability, and more lifelike skin details.

I have included a zip containing most of the caption files used in the dataset. Mine them for prompts.

What is this?

This model is purpose built to do one thing; generate photorealistic pornographic images of women in the style of a professional magazine photoshoot such as Playboy or Penthouse.

It has been designed to produce the most common poses found in such a photo shoot. Great effort and attention has been paid to lighting, color balance, saturation, and contrast.

Many hours have been spent attempting to produce a model that can generate accurate representations of female genitalia, as well as reasonable consistency with hands and feet, in-so-far as SD 1.5 allows.

The model works well with LORA and Textual Inversions, with some notable exceptions.

This model will not produce hardcore pornographic images that involve sex or penetration, and it does not play well with LORAs that facilitate it.

I am currently working on, and will release models in the near future that both play well with hardcore LORAs, and models that can produce such images on their own.

How was it made?

This model is the culmination of over a thousand man-hours of work. Over 2500 carefully cultivated images were manually cropped and captioned. This current release was first fine-tuned on Cyberdelia's Cyberrealistic Classic v2.1 as a base model using OneTrainer. 50 epochs with a 3e-6 training rate with EMA enabled. I used multi-res training at 512,576,640,704. It took a week on a 3060ti. This was the 5th (and hopefully final) training run.

The fine-tuned model was imperfect, so I merged at a 4.0 ratio with VanillaNudes from VanillaCoder, which was based on Cyberrealistic, which seemed to make it a good fit. VanillaCoder's attention to detail paid off, and this model is the result.

How to use

I have provided a wildcard text file (its in the download section, listed as training data) that contains poses and other prompts that can help lead the way. The word "ifap" was in every caption file in order to try to stabilize the specific style I was attempting to achieve. Sometimes it helps. The prompt "glamor photography" is also present in every caption, and this prompt is important to reproduce the style this model is intended to produce.

The basics are that this model can produce images of women of any race or skin color, sitting, standing, kneeling, squatting, laying, and on all fours. They can be naked, clothed, or in various states of undress. This model is meant to produce porn, so don't expect to not see some ladybits when you try to keep them fully clothed. The poses, for the most part, can be rotated front, side, or from behind. The model is capable of producing full body, partial, and close-up images of just about every part of a woman's anatomy.

It can handle higher resolutions, but becomes unstable, best results are keeping one dimension at 512 and not exceeding 768 on the other, then using highrezfix. Landscape images are more stable at 704 or lower.

The example images also provide a starting point for prompting. Keep in mind that I heavily used ADetailer in the samples.

Below is a list of all tags from the dataset that can have a meaningful impact.

these are in no particular order

ifap, glamor photography, playboy

posing for the camera, standing, kneeling, squatting, straddling, on all fours

from side, from front, from behind (looking back)

pale skin, shiny skin, tan skin, tan lines, tanlines, tan, bikini tan (pale skin is powerful)

breasts, flat chest, small, medium, large, huge, sagging

breasts apart

breasts hanging (on all fours)

sideboob, underboob

nipples (small, large)

pussy (best used alone for standing or kneeling, or all fours poses)

clitoris, clitoral hood (best used with sitting and laying poses)

spread legs, legs up, knees together, legs together

pubic hair (or female pubic hair, or both)

parted lips, smile, teeth, open mouth, grin

all natural eye colors and hair colors are available

hair over one eye, arms up, one breast out

underwear, panties, bikini, swimsuit, shirt, bra, undressing, thighhighs, high heels, sandals

dress (usually around waist), thong, underwear only, open clothes, no panties, socks, bikini top, sweater, lingerie, thigh strap

common jewelry such as necklaces and earring styles

piercing (navel, nipple, and pussy piercing)

masturbation, female masturbation, fingering, finger in pussy, finger to mouth

spread pussy, spreading her pussy, spreading pussy, rubbing pussy

ass, anus (if from the side or behind), urethra

indoors or outdoors, day or night (affect lighting)

laying or lying

mole ( on thigh, on breast, on ass, on body, on face, on stomach ect...)

freckles

blurry background, bokeh, depth of field, blurry

clothes lift, shirt lift, skirt lift, clothing aside, panties aside, panties to the side, panty pull, pants pull, no pants, bikini bottom aside, off shoulder, collarbone

topless, bottomless, barefoot, bra pull, underwear pulled down, pants, covering, sneakers, jeans, garter belt

makeup (not really needed, but sometimes adds something)

armpits, feet, fingernails, nail polish

presenting (ass up, hips out)

close-up (gets real close, works well with depth of field)

upper body (works well with "portrait photo")

photo background

holly randall (photography by)

cowboy shot

arm support, reclining, on stomach

tongue, grabbing own breast, soles (underside of feet)

hands on hips

ground vehicle, motor vehicle

from above, from below

Just because you can prompt things like hands and feet, this is still a 1.5 model, it's not recommended.

If you made it this far, whoa. Anyways, I hope you enjoy this. More models coming based the dataset.

A final note. I was reviewing my training data, and it appears I made a BIG mistake with my caption files. Over half of them had ended up on two lines, instead of one long line of tags. In OneTrainer, if you have the tags on separate lines, it treats them as separate captions. So, while it didn't really affect the quality of the images, it does greatly affect how closely the image follows a prompt. I've corrected this, and was going to do further runs anyways, so, the good news is, v2 will be much improved in several ways.

Thanks very much to VanillaCoder and Cyberdalia, this wouldn't be possible without their efforts.