Question about Wan 2.2 High Noise Model Training

I used the Wan 2.2 full model training script in this repository and encountered the following problems during training:

When I train the Low Noise Model, the training results can be well adapted to the original High Noise Model, and the inference results are all normal.

But when I started training the High Noise Model, things changed. I found that the loss at this time is about dozens of times higher than the low training. At the same time, when I combined the trained High with the original Low, the inference result would be a lot of noise, as shown in the figure below:

I'm not sure if anyone else has encountered similar problems. I guess training Lora should get the same conclusion as me. Is the training parameter configuration given by this repository the same as that used in the official training of Wan 2.2?

Start with the repository's Wan 2.2 full model training script and compare the Low Noise and High Noise training parameter configurations. Reproduce the reported loss and noisy inference result, then check whether the repository settings match the official Wan 2.2 training configuration; done means identifying the configuration or implementation difference, if one exists.