Diverse Male Nudity

TDLR here at the top

See below for greater (perhaps excessive) hints, detail, and discussion

• The primary trigger word is "penis". This main trigger was trained for average sized, cut, flaccid penises, and tends to produce such—though definitely not always. To push the model in other directions you can add one or more of the following additional triggers, preferably in this order: big/small, uncut, erect (e.g., "small uncut erect penis" or just "big penis").

• LoRA weight of 0.75–0.85 is the sweet spot in my experience, but I've gotten usable results anywhere from 0.6 to 1.0

• The LoRA functions and maintains its photorealistic output style with pretty much any Flux-compatible sampler/scheduler combo

• Flux tends to perform better for photorealism at lower guidance (2.1 to 2.8); however, being trained entirely on photos, this model can maintain a good photorealistic result up to 3.5 and a decent one to ~4.2

• If you're not getting prompt adherence for ethnicity, body type, or other features, try raising your guidance and possibly step count, and/or add repeat/synonymous descriptors to your prompt

• The model can be used for inpainting to improve penis detail, but works much better with plain Flux Dev, rather than the Flux inpainting models. When inpainting, expanding the mask to give additional context is helpful

Additional model limitations, suggestions, and technical notes:

Why I made it

Amid a growing selection of other penis LoRAs for Flux, I decided to create this LoRA both to try my hand at it and to attempt to expand the diversity of dudes and dicks that could be created. There are some great penis LoRAs out there, but I've found they tend to be biased toward fit men, white men, and circumsized penises with a similar shape.

Aiming for diversity has made the outputs more unpredictable than with other penis LoRAs, but has also, in my view, achieved more diverse outputs. I attempted to include a variety of ethnicities, ages, and body types in the training data. And I think the model's resulting abilities reflect this—even if the prompt adherence in this regard is not 100%. I'm admittedly biased, but this LoRA does seem better able to do non-athletic men and perhaps non-white men better than some other similar LoRAs.

Limitations and shortcomings

If anything, the model appears a bit biased toward bigger men and non-white men, which is kind of refreshing. Getting it to produce really thin guys, especially, can be kind of tough. If that's your thing, use a higher CFG and try a variety of phrases associated with that like thin, slight build, skinny, skinny arms and legs, etc. If you're trying to get blond guys and the results are looking a bit swarthy, specify blonde and/or some Scandinavian nationality.

Anatomical problems and visual artifacts both with the penis and the body still do occur. Lowering or raising the LoRA weight and/or changing the overall prompt can sometimes help. The more complex the overall prompt, the more likely problems become. For complex prompts, try putting the triggers at the beginning. The model does struggle sometimes with the penis hole, partly because it's not visible from all angles and because it looks quite different for cut/uncut. Short of re-rolling, this can be addressed with inpainting or possibly upscaling—or sometimes I just add one in post with manual painting or by repurposing a belly button or nostril.

The model does not handle butts or the anal area well. Sometimes you can get it produce a guy in profile view, but if you try for a view from the rear, be prepared for your man to have a butt penis (penis butt?). The model also does not really know masturbation, sex, or sexual positions.

Captioning choices

Although capable of creating a variety when it comes to cut/uncut, flaccid/erect, big/small, the model does not always take prompting direction well in these respects. To keep captioning/triggering simple and avoid needing to specify all aspects (i.e., always needing to use a full trigger phrase), I let the trigger "penis" on its own stand for flaccid, cut, average-sized penises; and I did not specifically train "cut" or "flaccid".

Perhaps I should have created completely separate keywords for each thing to improve adherence/outputs. But the other challenge is that sometimes uncut penises reveal the head or have the foreskin back and even erect penises can appear in different positions, so captioning is challenging regardless. I may attempt a retrain where I re-caption.

Nevertheless, using the additional triggers big/small, uncut, and/or erect still increases the likelihood of gens with those features. And repeating keywords also seems to increase the model taking direction (e.g., "large large uncut uncut penis"). Increasing CFG also (as you might expect) improves prompt following in these respects.

Training settings

The LoRA was trained on 140 images of nude men and close ups of crotches/penises. I trained for 35,000 steps (so far) at a low learning rate. This may seem like overkill, but I was still seeing improvements, even at 35k steps. Remarkably, the only captions I used were the triggers. To reiterate, no other description of the content of the training images was included in the captions. This just demonstrates how trainable Flux is and how good it is at "understanding" training images and concepts.

Odds and Ends

Unless you prompt for clothes (and even then still sometimes) the model will tend toward nudity whether or not you include the trigger. But the model is capable of SFW. The more you describe the clothes, the more likely the model will include them. Be warned, sometimes clothing will feature penis-shaped protrusions, which may be a plus for some people.