A LoRA trained around my own workflow and usage preferences.

Recently, I wanted to generate this particular motion again, but ran into the same old issues: either the penis part would deform easily, or the connection between penis and pussy would look completely locked and barely move. Getting a result that matched what I had in mind was surprisingly difficult.

I went back through my previous WAN2.2 videos, and honestly, there weren't many clips that met my standards. After filtering them down, I ended up with fewer than 10 usable samples and decided to train a version with them.

The results... improved some things, but they're still far from ideal.

Both the high and low noise were trained for 2000 steps, but after testing, I found that the 1000-step checkpoint performed relatively better, so the released version is the 1000-step one.

The dataset is still too small, so there may be a V2 version in the future if I can gather more data.