Finetune with Trajectory Condiiton

Hi, thanks for your great contributions! I have a question, when you initialized from pretrained weights, and you add the addition trajectory condiiton (as embedding added to time embedding), then through the AdaLN. I want to make sure that you trained **all** the parameters of the model? Since you changed the embedding input to the AdaLN. Have you tried other finetuned settings, such as freeze part of the parameters.

In my case, I freeze the pretrained weigths and only train the trajectory condiiton weights, it totally failed. And training all parameters leads to extremely low efficiency.

Any suggestions would be appreciate!

No file, test, or entry point is identified. Start by reviewing the training configuration and model code for pretrained-weight loading, trajectory-condition embeddings, and AdaLN, then compare freezing only condition weights with full-model training. Done would require a documented, reproducible recommendation for efficient finetuning.

Written by the indexing model from the issue text.