LoRA fine-tuning on Wan2.2 high noise and low noise models has little to no effect — anyone else seeing this?

Hi, thanks for the amazing work and for open-sourcing Wan2.2!

I've been experimenting with LoRA fine-tuning on the Wan2.2 A14B model and noticed some unexpected behavior. I'm wondering if anyone else has encountered similar issues, or has insights to share.

I used the official training script to fine-tune both the low noise and high noise models for a clothing/makeup transformation task, with a timestep boundary set to 0.875. The training process completed without issues.

However, after fine-tuning, the results were very poor — I tested using both the official inference workflow and KJ’s inference workflow, but the model failed to learn motion or makeup transfer effectively.

In contrast, when I used the same dataset and task to fine-tune Wan2.1, the results were much better, and the LoRA successfully learned the desired transformations.

Start by reproducing the report with the official Wan2.2 A14B training script, using the stated 0.875 timestep boundary for both noise models. Compare results through the official and KJ inference workflows against the Wan2.1 baseline; done means the cause of the ineffective LoRA learning is identified and the fix is validated on the clothing or makeup task.