"We performed additional training to improve the model's performance on high-resolution videos, avoiding issues like image corruption and grayscale."

Note: for I2V try to use big LoRA strength value such as 3.0

RTFM in Chinese