I am releasing 3 different checkpoints for this LoRA.

3000 is more flexible and arty, lets you get more interesting styles.

4000 is more strict about enforcing nudity and poses

5000 is right on the edge of being overtrained. It has some pretty warm output and a lot of contrast, even when you drop guidance down super low. That said, it can really create some amazing stuff, especially if you "overdrive it" at 1.20 weight (tho it cooks the output too like this)

toy around with weights and checkpoints, you're gonna get some magic and some freakshows. I'm sorry that not all of the nipples will look perfect yet, or that it will sometimes confuse bits and bobs. This model has very little nipple references and almost no genitalia references, so it's going to be some time and effort to bring this stuff back.

This version now has a specific keyword that you can use to boost output, scg-anatomy-female. The lora also has a much wider variety of dataset images that provide a bit more variety in output. It's really good at nipples now, and you can bias breast sizes {small, large, huge, round, voluminous, saggy, tiny, silicon implants} breasts, boobs, tits. It also has a marginal understanding of the concepts of "vagina, anus, pussy, asshole, pubic hair" and even can give you {small, large, dark, light, pale, black, pink, puffy, erect} areolas and nipples. it can do ages young adult (I didn't test younger and you shouldn't either) to elderly, obese to rail thin, muscular to pregnant to on her knees, bent over, (lots more, have fun trying it). The list goes on.

This version was trained on a little over 800 images, about 80% scraped vs 20% AI (AI was used to add racial and size variety, I was super super picky on what I used for inputs). All of the inputs were captioned first with Florence2, then hand processed to "remove the AI goofiness" and add explicit captioning to help add more natural keywording (just referring to a subject as "naked" and/or "nude" is usually enough, if you have a stubborn subject you can try describing more detailed features to bias the results).

This lora is not anywhere near "explicit ready" yet, and I'm not training it with any pony style commands, no easter eggs there, sorry. It's an honest first try now that I'm starting to get a feel for how flux trains with a cleaner female gender focused train. I recommend running the lora in the 0.4 to 0.8 range. While there is the occasional body horror hiccup, 9 times out of ten if I check the same seed without the lora the problem is on base flux as well, so 🤷🏻

Finally, I just want to mention that this took a lot of time and effort, especially the hand captioning work, as you really can't get ChatGPT to describe explicit material, and even the ablated VLMs suck at captioning naughty bits. I really appreciate it if you could give this a like and post some samples below. You get buzz, I get buzz, and hey, if you want to see more flux content (and I have A LOT more coming, not just nsfw stuff either) give me a follow, and if you're feeling really generous, click that coffee button and buy me a coffee, the tips I earn I put right back into my creations here. 💖💖💖