General futanari penis lora for Flux.2 Klein 9B. Trained on a synthetic dataset of photorealistic images.

Prompting: variants of "woman with a penis".

Important: The word "futanari" was not used in my captioning, so the model doesn't understand this.

Examples are all text-to-image.

T2I testing was done on the 9B distilled version (BF16), the following findings may differ for the base model.

- Trained on futanari images but it is flexible, it seems to work for men too but didn't really test that.
- Dataset contained solo futanari women with erect and semi-erect penises of various sizes, outputs lean more towards erect.
- Size of penis and testicles were mentioned in the dataset captioning so you can somewhat control size with prompting.
- Pubic hair was also captioned, but sometimes pubic hair gets added unprompted. If that is unwanted, try experimenting with different prompting or with "pubic hair" in negative or with NAG for the distilled version.
- Adding "veiny" seems to work if you want more pronounced veins.
- Should work for various poses and from various angles except from behind. Some poses/angles can be a bit janky.
- The dataset contained only solo subjects but it is flexible enough so it also works with multiple subjects, however don't expect most sexual interactions to work without other loras.
- Dataset leaned more towards light-skinned women so it will default to that, but different ethnicities and skincolors work when prompted.
- Defaults to a photorealstic style, you can add variants of "photo of..." to the prompt for more photorealism. Other styles still work, like anime, painting styles, etc.
Most of these things are shown in the example images.

The word "futanari" was not used in my captioning, the model seems to conflate this with various different things which affects results so I would avoid using it.

All example images have been generated with Flux.2 Klein 9B Distilled and gemma-8B both at BF16, with text only, no editing, and with the lora at strength 1.


This is a v1 of this lora and I will probably update it in the future as there is definitely room for improvement. I may make a 4B version too. You can follow my account to get notified of a new version release.

Please share your results and feedback, this helps improving future versions!
Keep CivitAI's TOS and common sense in mind when sharing results here.

Leave a like if you enjoy the lora :)

Ko-Fi: ko-fi.com/degenerator123

Added training info:
Dataset of 200 synthetic photorealistic images with captions in natural language, training res 512, 768, 1024. At around 3k steps it was working well. The dataset used is a regular dataset of images with captions, not an editing dataset.