TLDR: Can do a diversity of penis types/sizes, ages, body types, and ethnicities (see triggers). Higher resolution / closer shot = better results. Does not require uncensor LoRA; but increasing weight helps prevent refusal.Trigger list (with explanations where needed):<ul><li>Penis</li><li>Flaccid</li><li>Stiff (engorged, but not pointing up; this one is very unreliable)</li><li>Erect</li><li>Uncut (uncircumcised)</li><li>Retracted (uncircumcised but with head fully/partially exposed)</li><li>Circumcised</li><li>Large</li><li>Small</li><li>POV view</li><li>Nude (untrained but works well and decreases refusal)</li></ul><h3 id="whyhow-did-i-make-it">Why/How did I make it?</h3>(most of you can skip this)This LoRA is the successor to my Diverse Male Nudity LoRA for Flux 1 and 2. I opted to make this a new LoRA rather than an update because I wanted to update the name (I think people search more for "penis") and because I changed the tagging to hopefully make the LoRA more flexible and responsive. The file size is large because I used linear/alpha 64, which is what enabled the LoRA to cover a variety of penis types and make them fairly promptable.There are a lot of serviceable penis LoRAs out there, but I find each one tends to cover one particular "type" of penis, or even just one very specific penis. I've also found they tend to be biased toward fit men, white men, and circumcised penises. I wanted a LoRA that could do a variety of looks, so I trained on a diverse dataset of over 200 specifically captioned images.The LoRA was trained entirely on real photos. It can still do other styles very well, if prompted, but adherence/coherence may suffer a bit. (Pro-tip: To get better much more accurate styles, put your style language at the beginning e.g. "A cubist Picasso painting of...")I did a lot of testing to try to find the best number of steps, but with fairly high seed-to-seed variety, it was difficult to be sure which offered the highest adherence/quality. (The fact that it came out to 6960 was pure coincidence). I think this one was the best balance, but I may train more steps in the future.<h3 id="what-the-model-does-wellpoorly">What the model does well/poorly</h3>The model handles a diversity of ethnicities and body types extremely well. That said, it does seem to associate some penis types with certain body types or ethnicities more than others, so you may need to workshop your prompts. (I tried to get diverse penises across diverse men, but I could only gather/tag so many penises.) It can do different sizes and even extreme sizes ("very, very large",) but adherence gets worse the more extreme you go. You can try and sometimes get "results" with "penis is 3feet long" and weight cranked to &gt;1.5, but this was not in the dataset and body horror this way lies.Please note that some combinations work better than others. For example, it does better with uncut flaccid/small than uncut erect/large—probably because erect penises, which are also the larger state, often have foreskin retracted. The converse is also true, with circumcised working better for large and erect. Unfortunately, "stiff" is a bit of a weird trigger and the most likely to screw things up.You can also sometimes specify "penis pointing [up/down]" to get closer to the look you want. Overall, increasing the base generation resolution improves adherence and appearance somewhat. But in the end, it's still not perfect, so don't be afraid to re-roll as necessary.The prompting can be a little finicky, but this LoRA can produce excellent first-person POV views. I know it's redundantly worded, but add "POV view" at the beginning of the prompt and then be sure to word your prompt such that it does not sound like looking at a person. (e.g., instead of "a male sitting on a chair in a bedroom" "looking down at a male chest, stomach, and penis. The setting is a chair in a bedroom").<h3 id="recommended-settingsstrategies">Recommended settings/strategies</h3>I strongly recommend at least 8 steps for optimal detail/realism. A weight of 1.0–1.25 seems to work well, but increasing the weight to ~1.5 can sometimes improve adherence/coherence and help with refusal (I think this may mean I could do even more training). In my experience, Euler/Simple are the best sampler/scheduler combo, but others that work with Krea 2 also work. Linear quadratic, for example, seems to offer more saturated colors.You don't need to use an uncensoring LoRA, but you can, if you want. It does serve to further reduce refusal and can sometimes increase prompt adherence. But be warned, there are a lot of different uncensoring LoRAs out there with different optimal weights, and combining them with this LoRA can get weird fast, especially at higher weights.<h3 id="model-bonus-abilities">Model bonus abilities</h3>Fascinatingly, you can also use this LoRA to sometimes improve SFW digital/smartphone camera realism and prompt adherence as well. The first part of this makes sense because it was trained on photos from digital cameras and smartphones. The second part is more mystifying. My best explanation is that the LoRA dampens the refusal layers, which others have discovered negatively impact prompt adherence in general. Anyway, the LoRA seems to not randomly insert men or penises unprompted, but be careful and scrub your metadata :PThe LoRA dramatically improves prompt adherence with respect to body type and facial features. It also appears to improve image variety between seeds. Additionally, with use of the untrained trigger (nude) it will also create images of nude women. I think these represent additional instances of the LoRA muting the refusal layers.<h3 id="addendum">Addendum</h3>This model is also very happy to create penis shaped objects and add penises to things. It will also readily create female-presenting people with penises. So get creative and have fun!<h3 id="disclaimer">Disclaimer</h3>Do not use this model for any unethical and/or illegal purposes, as determined by the jurisdiction you live in. You are solely responsible for whatever you create using this LoRA.

Files with SHA256: 539655b1d4669ec0fc9a764b63f031800b1404ac1b16b5f5cff1d7d5d7f62c90