Concept

This was an experiment to see what I could get WAN to do with very little input, the results came out decent so I thought I'd share. I wanted to dress a character in a specific outfit from VAM (Noc_this maid outfit) and was really struggling to get WAN to understand my prompt. So I booted up VAM, took 17 screenshots of a character with the outfit on from various angles and different boob sizes, auto-captioned them using Joycaption, punted it all into diffusion pipe and a few hours later got a working model. The results are reasonably accurate so decided to share, so everyone can have their own big tiddy goth maid.

Usage

Should work in both T2V and I2V. Prompt is simply:

Base prompt T2V is:

A woman wearing a exposed maid outfit. She wears a black, long-sleeved leotard with a high neckline and a white maid apron that exposes her breasts.The apron has a frilled edge and ties at the back. The leotard is very high-waisted. 

For I2V assuming you're using the outfit change lora:

The woman has a sudden outfit change into an exposed maid outfit. She wears a black, long-sleeved leotard with a high neckline and a white maid apron that exposes her breasts.The apron has a frilled edge and ties at the back. The leotard is very high-waisted. 

Thats it, enjoy.

Training

As mentioned, 17 images with 2 repeats trained at 0.6mp on high and 1mp on low. Batch 14 and 8 respectively. Only trained it for 25 epochs on both low and high. LR was quite high as I was impatient at 3e-4. Trained on diffusion pipe on 4x3090s