Evolved futas!

<click "show more" to see full description for usage suggestions>

Trained on a photorealistic checkpoint: PonyRealism v2.2 (linked below) but seems to work on other Pony based checkpoints too. Play around with the strength to see what works for you, should work pretty well between 0.7 and 1. At higher strength it will generate triple layered tips more consistently, with lower strength it can also generate double layered tips. Adding the trigger word in prompt is not strictly necessary, seems to work without it too, but may be more consistent with the trigger word included.

Trigger word: t3p

This is my first shared LoRA, it took quite a while to create because I had to make the entire training dataset myself since no real images of this concept exist.

Trained and tested to work with:
- Different angles: from side, from below, close-up, from behind, etc
- Different aspect ratios
- Different poses and positions
- Multiple characters
- Different penis sizes
- Erect and flaccid
- Different skincolors
- Combining with other LoRAs. Works well with RealismEnhancer LoRA set at a lower strength (around 0.3-0.5, linked below)
- Other checkpoints. May have an effect on style, experiment with different strengths
- Pubic hair, when prompted. Does not generate pubic hair unprompted
- Not generating underage characters randomly, training dataset included no underage characters
- Veins, adding "veiny" to positive or negative works

The above list has been tested on the PonyRealism v2.2 checkpoint, results may vary between different checkpoints. This LoRA is trained on and trained for futa. With very limited testing (aka 1 image) it seems to work on male characters too if that's what you are interested in, no promises there though.
The LoRa is trained on futa with penis+testicles, but with the way I tagged the dataset it should be compatible with different types of futa (no testicles, with vagina, etc.)
Sometimes works on other types of penises, may require some prompt experimentation and a higher LoRA strength. This is hit-or-miss, so roll the dice a few times to see what works

Known issues / room for improvement:
- Seems to favor the color blue when outfit and background is not prompted
- Sometimes puts a hat on characters when outfit is not prompted. This won't happen where it wouldn't make sense, like when laying in bed, showering etc, but may happen when standing or outdoors. If this becomes an issue add "hat" to negative
- Certain background elements may show up repeatedly across generations when background/setting is not prompted

These issues are probably due to the relatively small dataset and will get fixed in a future V2 by expanding/improving the dataset and tweaking some training settings, however the core concept should work well in this version.
You can follow my account to get notified of a new version release

Example images
CivitAI seems unable to read metadata from my images, I will look into that for the next time I upload. Example images have been generated using PonyRealismV2.2 at 35 steps, CFG 6.5, DPM++ 2M SDE Karras, some with the RealismEnhancer LoRA at 0.3-0.5, no editing/fixes/upscaling and no other LoRAs.
Basic prompt setup I used for example images:
- Positive: score_9, score_8_up, score_7_up, t3p, futanari, ...
- Negative: score_4, score_5, score_6, muscular (muscular in negative is specific to the checkpoint and my taste, you can add whatever you want here)
I kept the example images pretty tame, but as stated before it should be quite flexible.

Since this is my first shared LoRA any feedback is appreciated, also share your generations here so others can see what works! I assume this will probably remain a pretty niche LoRA, let me know if you are interested in this or similar concepts.

I prefer local generation/training with more control, unfortunately my setup is getting quite old and slow. I'm currently saving up to upgrade my graphics card so I can train and test LoRAs faster, I'm definitely not expecting anything but should you feel inclined to support that, I set up a donation link

Programs used:
Training LoRA: OneTrainer
Editing images for training dataset: Krita with ComfyUI plugin
Generating images and testing: ComfyUI

Checkpoint (not my own): PonyRealism(V2.2)
Optional LoRA (not my own): PonyRealismEnhancer(V1.0)

Disclaimer: I do not condone the use of my models for generating imagery of underage characters, existing people such as celebrities, depictions of potentially illegal acts or anything else that can generally be considered problematic.
Do not upload these types of images here.