This model will generate THE best penis that I've seen for a text to image model at the time of this writing (June 2026). You can prompt vein amount, size, and other anatomy as done in my showcase images.

The body and pubic hair is the best I've seen and it even handles dark wet hair, which is usually a fail (I'm talking to you Qwen Image and your squiggly lines for leg hair)

Please tip if you find my stuff useful and want more - it does take significant time to keep with all the latest models.

Give this combination of sampler/scheduler a try: exp_heun_2_x0_sde + linear_quadratic

v2

Trained for 10k steps total of 10 hours on my 5090, same dataset but focused training on the low noise to improve details.

v1

It only needed 4.5 hours to train v1 on my 5090 against the raw model and my 200 image dataset, therefore it can be used with both the turbo model (all of my showcase images are turbo, 8 steps) or the raw model. I didn't get a single backwards dickhead, and only a couple detached penises that were easily fixed by modifying the prompt. My 20-image showcase is only cherry picked to exclude the images that didn't show D in them. It is that good!

It knows many A-list celebs already without images or loras, but their safety feature added more clothing when I tried to use my lora with them. I put a link in my Collection to the Krea2FilterBypass lora which you should use if mine doesn't give you any results.