[Question] Inquiries regarding training details and implementation of Wan-Animate

[Reed-yang]

Hello,

I am currently working on implementing the training and fine-tuning pipeline for the Wan-Animate model using this codebase, but I have encountered some difficulties.

Could you please provide some clarification on the following points?

1. Loss Formulation: Is there any difference in the loss function (or training objective) between Wan-Animate and Wan-2.2-I2V?
2. Training Verification: Has the Wan-Animate model been fully trained or verified using this codebase?
3. Data: Could you share details about the training data used for Wan-Animate?

Thank you for your help!

[bubbliiiing]

1、The loss objective of FLOW is the same, whether it's I2V or T2V, they share the same objective.
2、I built a small dataset (about 10 samples) to verify the accuracy of the loss and whether training would collapse. I haven't trained on a large scale yet. Do you have any dataset recommendations for me to test?