!WIP MODEL!

Does what it says on the tin, nudges images towards missionary POV-ish view (non-pov soon). Handles dongs decently well without additional LoRA (dong lora soon), which is neat. Recommended weight around 0.5-0.8 to avoid watermarks.

Versions

All the versions are basically the same, but have different quirks. I recommend trying them all out.

Missionary - Base model

Alt - Base model with OUT blocks slightly pruned in an attempt to fix watermarks, trends more towards vagina-holes sadly. Will probably fix later

Half - Midway checkpoint, I like it more for certain prompts

Current issues

• Watermarks: Dataset contained a lot of images w/ watermarks, LoRA will occasionally output, will rebuild when I'm less lazy

• Mysterious left-corner man: For whatever reason it really latched on to 3 images of men in the corner of the frame, idk how to fix it just drop the weight

• "riding" pose: Model quirk, I can't really do anything about this due to the poses being similar

• Condom: There is 1 (o n e) image of a dong with a condom in the dataset (now nuked)

Prompt advice

I recommend avoiding plurals as it can get messy pretty quick. Use man instead of men, etc. Negative weighting hands also helps with random hands appearing, but it'll always be flawed.

LoRA need something to latch onto, the model has no text encoder training so you have to give it something to go off of. I recommend either throwing in sex or anal, either work, 'lying' also helps, and I've gotten good results with "solo". Also recommend leaving out penis unless you're doing solo-heavy images.

Feel free to message me on discord @fleshtexture if you'd like help