Help - Training WAN 2.1 or 2.2 Lora Models #662
Description
When I train Hunyuan lora models they always look really good, but when I train WAN 2.1 or 2.2 models to use with T2V or I2V using the same datasets and configs (tailored to train with WAN models) I get very poor results
Often the faces and bodies are distorted or noisy....
Would a kind person please share their technique with me to train good quality WAN 2.1 and 2.2 lora models of people please?
Can you include a config to train these models using musubi-tuner so i can compare against my settings to see where i'm going wrong?
Also how long you train the model for and at what resolution or training speed please? any help greatly appreciated 👍