Most of the time, the regenderings go fairly smoothly. With some of the modern LoRAs like SNOFS or DickTator, it can be as easy as “change pronouns, add a LoRA”. But every once in a while, there is a real challenge. In this case, I think it has a lot to do with the LoRA. It is doing a lot of work in the generation. In particular, it has a very specific view about lower torso anatomy: at least two holes, maybe more. And definitely no poles.

So, without further ado, let’s talk about the original:

Mostly nude female bent over with jeweled buttplug. By the side of a pool with palms in the background
How much of this is the model, and how much of it is the LoRA

This is a combination of Flux.2 Klein 9B, and a LoRA, “Real Anatomy Ultimate Intimacy”. The LoRA’s “official” trigger words are pretty simple (“plug”), but both the suggested prompting and the examples are far more detailed. Here’s the original prompt:

plug, A low-angle rear close-up shot captures a woman bending over by a pool. She has light skin tone, long black hair, facial tattoos on cheeks and neck, visible buttocks and genitalia. Her pose features her spine arched, hips posteriorly tilted ~60 degrees, knees flexed at ~90 degrees with weight on bilateral knees. The action is bending over while having red jewel butt plug into the anus. She wears black fishnet stockings with diamond pattern, pink bikini bottoms pulled down to thighs, and black high-heeled platform shoes with studded straps. The setting includes a poolside area with palm trees in the background under a clear blue sky. Lighting is bright daylight from overhead creating soft shadows on skin and clothing.

The prompt is amazingly detailed, all the way down to the exact angles the woman takes on as she is bending over. This seems to be the direction the entire field is heading, allowing for ever more detailed levels of description. Now, the question here is how much of this image is from underlying model knowledge, and how much is from the LoRA.

Out of number of generations and iterations, this is the best version I got:

alt text

And this is, well, good, but not great. His body mostly reads as masculine, but his dick is not at all visible, and his balls are pretty damn tiny. Unfortunately, this is really the best I could get for this combination of prompt and LoRA.

The failures were pretty bad. The “best” was just missing his arms:

A very disarming image

Some were fairly disturbing: it’s not just the random neck-dick (or the butt dick), but the way the jewels are physically bonded to the model. :

Almost nothing about this image is good

I’ve touched on this before, but all of these models really, really want to generate vulvas everywhere1.

Tell me how much you love pussy?

Now, the challenge with this LoRA is that it was clearly trained only with women. The LoRA is layered on top of Flux.2’s exiting biases, which is lightly female. (I don’t have a solid number on Flux.2’s bias, nor a good idea about how to get it right now. But something to think about here.) Together, these two create a fairly high local maxima (minima?) that is hard to escape from. And, so, we get a LoRA that really can only create women.

I could probably rescue this through img2img type generation, but I feel like that defeats the purpose of Regendering Monday. I am lazy, and I don’t want to work any harder than I have to. Really, pronoun replacement and maybe changing how breasts are talked about is about all I am ready for.

  1. Wasn’t there a William S. Burroughs story about people who could move their genitalia around on their body, maybe even able to move to them to other people’s bodies? My memory is that these were a bit dated in the early 90s; they would practically ancient by now. ↩