The transferability of watermark

[xbtzpx]

Hi, I trained the encoder and decoder myself and then embedded the identical fingerprint with the encoder (set random seeds).
The watermarked data set is fed into the CUT model for training, but the generated image is sent to the decoder, and the decoded watermark is almost completely different from the one I embedded.
I want to know the details of your network training with watermarked data. Can you tell me more about it?
For the CUT model, I trained 400 epochs.
Can you answer my question?