Declaimer:
Not sure if I should even be uploading this, because this Lora is not quite end-user quality.
But I’m using it to make images, and just perhaps it will help someone else as well.
So, what is this?
It’s a Realistic Size Flaccid Male Genital Lora (at least my best attempt at it so far). You know, the kind you see on non-pornographic male nude imagery.
Does it work?
Like I said, it’s not plug and play, as I, ideally, would like it to be. There is a workflow one needs to follow to actually inpaint modestly sized genitals.
It takes about as much inpaints as the rest of the image 😀 Buckle up.
(Apologies to anyone on very low specs – this will hardly be useful)
Wat do?
1. Generate the image you want genitals on, with any checkpoint you like (I’ll test some as I go, but I can’t try them all; models that already can generate male genitals, at least to some degree, are a good bet).
2. Go directly to inpaint (before any kind of Upscaling or Hires fixing).
Mask abdominal area, along with any undesired genitals generated by the checkpoint.
Inpaint Masked. Fill/Latent Noise. Whole Picture. Set Size to your full image size. Denoise at 0.75-1.
(this is important, because you want to inpaint with perspective to entire image, otherwise, genitals will be too large)
Add Lora to prompt at 1.25 or higher Weight (and nothing else, maybe ‘nude man’ at most, any other penisy keywords will only enlarge the genitals).
Generate a few batches. Right now, you’re looking for a good believable outline for small flaccid genitals (just a vague base you will add detail to later).
3. Now do your upscaling (without the Lora), whichever method you prefer, but keep in mind that you want the genital outline to stay intact (it’s possible to do Step 2, after upscaling, but then you’ll be inpainting huge images – not ideal).
4. Go back to inpainting and inpaint genitals locally, at around 0.35-0.5 denoise until you get some amount of detail. This step varies a lot per checkpoint you’re using. If it already can produce genitals you like – inpaint without Lora. If it can’t – use Lora with Weight 1-0.5. Prompting for ‘penis, flaccid, uncut’ at this stage helps. In the end, it’s a limp dick, don’t expect SD to have learned anything pretty other than ‘skin folds all over’.
Wew, this easy?
I’ve done limited testing so far and have a few things to advise.
Models with HUGE genitals are pretty stubborn. Don’t use any genital related keywords, sometimes even ‘nude’ will increase genital size.
If initial genitals are too big. Try to inpaint something smaller first, then mask a smaller zone and inpaint further.
Models that favor female genitalia will produce something akin to ‘penis coming out of vagina’ a lot. If you see male genitals entirely surrounded by skin folds – add “female, woman, girl, 1girl, vagina, pussy” and whatever else comes to mind to Negative Prompt.
I’m not 100% on this, but try not to use negative embeddings when doing Step 2. Basic Negative Prompt, doesn’t seem to hurt.
I will not do any editing for previews (outside of nice little frame), but, honestly, it’s a lot easier to paste in Step 2 genitals after Upscaling, rather than try to upscale without affecting them.
This was trained on photos, but I can’t promise photorealism, because I don’t normally generate that sort of imagery.
Lora was trained on ewolf’s OhmenLifelike. Lifelike versions performed MASSIVELY better than same dataset trained on SD 1.5. I will maybe post the SD 1.5 version as well. Just to make it even more confusing.
Above point also implies that this lora works best on all derivatives of OhmenLifelike, supposedly.
Thanks, but no thanks.
Honestly, I don’t know what’s going on with Stable Diffusion’s 1.5 penis concept, but that thing is all over the place. It will learn anything and everything else that’s in the image, just not the genitals.
Perhaps this workflow works even better with one of the other penis Loras on CivitAI, I don’t know, I didn’t try (this Lora had exclusively uncut tiny subjects in training data, except for close-ups). If you feel like it. Try it.