Georgefwt
commented
2 months ago This is an interesting problem. It can be observed that the training input for the diffusion model (including LoRA) contains watermarked samples, while the input for the frozen diffusion model consists of clean samples. It is through this distinction that the watermark information is introduced. A detailed proof can be found in Appendix E of our paper.
Hi! I've read the part of ppft stage in paper and code. I wonder why the watermark remains preserved during the denoising phase if we ensure the predicted noise $\bar{\epsilon}_{pred}$ and $$\epsilon{pred}$$ are similar. In other words, how does the training methodology ensure the preservation of the watermark, despite the absence of any watermark-related loss during the PPFT phase?