Concept

An experimental small/micro penis LoRA for MiniMax H3 i trained a month or so ago and hadn't gotten around to releasing. Minimax isn't great at cocks to start with but pretty much every mix or cock lora generates giant erect circumcised cocks. Even the flaccid penis lora generates pretty large cocks. The original goal was just to create a better cock lora but trimmed the scope down to small cocks only as that was the major missing feature. Foreskin is supported but its super hit and miss, i've tried captioning it a whole bunch of ways and none seem to give the result i want but it can work sometimes.

This couples well with the softsuck lora hence the timing. The lora also couples pretty well with @diogod's excellent male anatomy lora (https://civarchive.com/models/2839513/male-anatomy-ass-penis-testicles-and-glans-h3). The second example video uses both loras together to get a better overall result.

Usage

Should work in both T2V and I2V as an anatomy anchor. It works much like a slider lora where the stronger you put the weight the smaller the cock. I'd recommend no lower than 0.6 and no higher than 1.2.

Prompt

So first thing to note is i somehow managed to misspell flaccid and the captions all contain flacid instead. so make sure you use flacid when describing the state of the penis. Some sample prompts to get you started, though Minimax really does need longer prompts than just these. Check the samples for good full prompts.

Size (small/micro)

 a small/micro penis.

State (flacid/erect)

 a small flacid/erect penis.

Foreskin(intact/circumcised)

 a small intact/circumcised penis.

Foreskin position (retracted / extended-covering / partially retracted)

 a small intact penis with the foreskin retracted / extended / partially retracted.

I do tend to find that it tends to do darker shafts by default. I think this is largely due to the lack of visible shaft in most of the input images. You can try and prompt the shaft color as being pale or dark but it has mixed results.

Training

The model was trained on 100 images at 2mp for about 10000 steps. Captioning was automated using Qwen3 VL 32B heretic. Training was run using diffusion pipe on 3x3090s. I used the diffusion pipe image only training adapter to avoid images slowing motion and it seems to have worked well.