Trained on 550 photographs with natural language captions of men of all shapes and sizes from the front and the back, this LoRA allows the full male anatomy to be represented on the Flux.1 dev model by Black Forest Labs. This is intended for the Dev model, I can't make any guarantees as to Schnell. I use Forge with the flux1-dev-Q4_0.gguf variant and Diffusion in Low Bits set to "Automatic (fp16 LoRA)". I can't help with other GUI's or model variants, they're out of my reach regarding resources. If you have problems, try to recreate one of the sample images as closely as possible to test your setup. If you figure out ideal settings, I'd love to hear from you in the comments. As well, after a fair amount of testing, I highly recommend a CFG of &gt;=2 for more successful generations.In addition, you'll likely have trouble stacking LoRAs with this. If you're hellbent on it, I'd recommend following the Article by rainfiend linked in the resources below and giving that a try. Both solo and group images including a handful of close-ups were used during training. In terms of tagging, both "circumcised" and "uncircumcised" penises were used in the captions as well as details regarding the "foreskin" but I don't believe it learned the difference so I usually don't bother including such details in my prompt. The state was also included such as "erect" or "flaccid", sometimes "turgid" or "half-erect". From behind, "anus" and "asshole" were used in equal turns. I don't think you need to mention it unless you want an explicit focus on that part of the anatomy. Some other tags of interest may include "pov" (e.g. "a pov photograph from the viewer's perspective of") or mention of the viewer's limbs such as their penis, hand, thighs, etc. A limited amount of pose data was included. Don't expect hardcore pornography from this LoRA unless you're a very clever prompt smith. It's intended to serve as a somewhat flexible male nudity hotfix. not a revolution in porn and complex, multi-individual prompting.

Project Men - FLUX LoRA

Some training details if you're interested:<ul><li>Utilized the CivitAI trainer, cost was around 8k Buzz.</li><li>70 Epochs with 1 Repeat and a Batch Size of 4.</li><li>This ended up being 9,625 steps.</li><li>A full resolution of 1024 with Bucketing was used, and CivitAI automatically resized images larger than 2048.</li><li>Unet LR of 0.00010</li><li>Scheduler was constant</li><li>No min SNR gamma or noise offset</li><li>Adafactor Optimizer</li><li>128 network dim and network alpha</li></ul>There's no chance in the world these were the absolute best settings but they finally got me the results I was looking for.

Files with SHA256: bb49df31244639ccf44b89b8924acd734698e96a7dc6d508d2dd873cb7781322