This is an experimental piece. I just wanted to test whether some poses that defy normal human body structure could be effectively trained into a Lora, and the results have been very satisfying! This motivates me to expand the dataset
I only used six 33-frame videos to train it, and three of those videos were generated by me using wan2.1 vace. This was challenging because DWpose couldn't recognize these unusual movements!

I only tried testing it with some simple prompts; you can try using complex prompts and hopefully it will perform well too.

Tested at 480x848 and 720x1280 resolutions, not sure how it performs at other resolutions.

The girl is wearing a blue bikini and black high heels, 
showcasing the "sanzhe" pose on a dimly lit stage, 
holding the tips of her shoes with both hands, her face showing pain

I've only tested it using wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors with 4-step sampling

I don't know how this Lora performs without acceleration.