search

bsmhmmlf / Gaussian-Shading

[CVPR 2024] Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models
MIT License
46 stars 3 forks source link

About generate identical artwork and watermarked images #4

Open Jayamoon59 opened 2 months ago

Jayamoon59 commented 2 months ago

Hello, I've been learning your code lately, and now I can generate a watermark image randomly according to the prompts, but I want to generate the original image and the watermark image of the exact same image like the experiment in the paper? What should I do? I hope you are available to help me with this simple issue, thanks！

bsmhmmlf commented 2 months ago

Hello. What we are trying to say in the paper is that gaussian shading can generate the same image as SD as long as the input latent representation is identical, without any effect on pixels and semantics, but at this point it appears from gaussian shading that the image contains watermark information. So if you want to get the same image as SD, you still have to keep the input latent representations the same.

Hello, I've been learning your code lately, and now I can generate a watermark image randomly according to the prompts, but I want to generate the original image and the watermark image of the exact same image like the experiment in the paper? What should I do? I hope you are available to help me with this simple issue, thanks！

Jayamoon59 commented 2 months ago

Thank you for your reply. I understand that if you want to generate the same graph you need to input the same underlying vector. But what I don't understand is why the inclusion of a watermark doesn't affect the underlying vector? And I did not get the correct result graph (555><) when the code generated the original image, I hope to get your advice