REPRINT
This lora is meant to be used as a helper for Wan's under-trained genitals. It was trained with the T2V model and diffusion-pipe on a dataset of 50 images and 30 videos with the faces cropped out for 20k steps, it's quite over trained but works well with I2V. You may need to reduce strength with T2V.
The dataset uses 2 models, one with an innie ***** and one with an outie, for both consistency and variety.
The Lora was trained on nude poses in the following positions:
Standing front and rear.
Sitting/Lying with legs spread open.
Bending over or on all fours from behind.
So it should help with genitals in any of those positions. The trigger words are Innie vagina or Outie vagina, depending on your preference.
I've planned this lora for quite a while with a very carefully selected dataset, and I'm excited to release it here to the community. Please upload your creations and tag this model so they are featured in the gallery! I'll buzz and feature the best gens.
For other trainers interested in the details, I set up this lora so that it would train on high res images and low res videos at the same time, the images for the detail and the videos for the movement. This seems like an effective technique for Wan as it's good with motion so the images don't degrade the lora much and provide great detail where needed. Using low res videos also greatly speeds up the training. Learning rate was 2e-5.