about training Concat-ID-Wan2.1

[RongPiKing]

When training Concat-ID-Wan2.1, if some of the later frames in the generated video are simply the input image condition, how can this issue be resolved?

[yongzhong-ruc]

You can either train Concat-ID-Wan2.1-Adaln to solve this problem or collect more data to train Concat-ID-Wan2.1. I recommend Concat-ID-Wan2.1-Adaln.

[yongzhong-ruc]

I only released the code for training Concat-ID-Wan-AdaLN, so I don't understand what you mean by "training Concat-ID-Wan2.1". Also, please check whether your training and inference setups are consistent.

[yongzhong-ruc]

Additionally, I suggest using the weights we have released for initialization by setting the training parameter '--resume-dit_path={your_dir}/Concat-ID-Wan/second_stage_adaln.pt'.

[RongPiKing]

Thank you for your reply.