This is an unofficial implementation of the Paper by Kejiang Chen et.al. on Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models
May I ask why the generated images before watermarking and after watermarking are different? Since the results in the paper indicate the watermarked image and the original image are almost the same.
Hello, first of all thanks for your great work on implementing Gaussian shading!
Could you explain how 0.18215 comes from in your method img_to_latents, does that mean scaling the latent after vae encoding? why it has to be this number?
Thanks in Advance
Different from the result reported in the paper of random crop, it turns out that the bit accuracy of crop with ratio 0.9 achieves about 0.6, and ratio 0.7 or less achieves about 0.5.
Did you also achieve this result?
I think this result might be reasonable because the whole latent is divided into chunks of 256 bits, and then voting is done on each position. However, the positions themselves may not correspond to the correct starting point.
Maybe for extraction we need to start from different points and choose the best result? I'm not sure if my analysis is correct.