<h1 id="heading-307">iFAP - X</h1><h2 id="heading-2701">What is this?</h2>This model is purpose built to produce explicit pornographic images of a man and woman performing sexual intercourse.The motivation behind this model was the ability to portray a moderate range of sexual positions through the use of wildcards instead of relying on LORAs with a focus on quality, consistency, and reliability.This BETA release is as close as I have been able to come to reaching that goal. It is not perfect, but it is close to being "good enough", capable of meeting the initial criteria if you prompt correctly and allow for a generous number of generations in order to meet your needs.I'm not done, but I figured I'd put this out there. The feedback I received from my last release was invaluable in improving quality. More is coming. I have several different fine-tunes complete based on several different base models. This is just a test-release to gauge interest and receive feedback.I have tested this model with a wide array of LORAs, and it's a mixed bag. It is too complex a topic to provide much guidance or recommendation. Try things. See what works. Some things conflict horribly, some things work really well.<h2 id="heading-2702">How was it made?</h2>This model started as a fine-tune of SD1.5 using a custom dataset of over 1500 carefully cropped and captioned images. This beta release is a merge between this custom fine-tune and Pornrealistic_v10 at a 4.0 ratio.It was trained on images from professional pornographic productions, it doesn't do well with "amateur style".<h2 id="heading-2703">How to use</h2>This model was trained for a handful of very specific sex positions and camera angles. While the merged model allows for some flexibility, attempting to prompt outside of the limitations will result in unreliable and poor quality output.This model does the following well enough that I decided to release this BETA:Doggystyle position, filmed from the side (not from the front or back). Works best in landscape orientation with a maximum length of 704px. It is the least reliable, it takes a good number of generations to get reasonable results sometimes. The use of a LORA at low strength can improve output, but does defeat the purpose of the model.Missionary position, both POV, and from the side. Side view works best in landscape orientation, POV works best in portrait. Isn't perfect, the use of adetailer is encouraged.Cowgirl position, from the side, and from behind. Works well at a variety of dimensions.Reverse cowgirl position. Clearly, SD1.5 has this in the dataset, and models heavily favor it. It took a lot of work to get this position to not overtrain while allowing for a balance with the other positions. RCP produces the highest quality images at any orientation, and is the most flexible for pixel count.If you run this model unprompted, or with a limited prompt, you're likely to get a lot of RCP images. Don't try to use this model for non-NSFW images. You'll get fired. It's over-trained. It does porn. It does it well. That's all it does.This model can do vaginal and anal. It can do white people, brown people, and black people.There is NO ORAL SEX trained in this model. Sorry BJ lovers. Don't bother, even with LORAs, it won't work well.With LORAs this model produces pretty good threesome images, especially DP. I have no idea why, it's not in the dataset, but whatever.Best results are found keeping one dimension at 512 and not exceeding 768 for portrait, or 704 for landscape orientation.The example images provide a starting point for prompting.I have uploaded both the lion's share of the caption txt files from the dataset to allow for more fine-grained prompting. There's a lot to mine in there. I have also uploaded a wildcard file for use with dynamic prompt extensions, which was the entire point of creating this model.

iFAP-X Beta 2

Beta 2 - same dataset, different learning rate. 5e-6 vs 4e-6, a little overbaked, but i find myself using it a lot, so there's something about it that looks better to me than beta 1. Try it yourself, see what you think.

Files with SHA256: f9b4b77c35f4743da653e106162d029df18e075ef4137ef8cf9afde4ef4a7724