You have two load diffusion model nodes that load your High and low models, respectively. Then in the model paths of each, before sampling, load these loras. Literally saying the word "twerk" or any mention of hips or whatever, this lora tends to activate. The first version was very strong.

https://github.com/kijai/ComfyUI-WanVideoWrapper/blob/main/example_workflows/wanvideo_T2V_example_03.json

The folder this file is in contains so many examples. All quite clean and mostly vanilla.

Even though it's a T2V base, you can still use it with I2V models.