About the visualization of the ring pattern in Fourier domain

[qilong-zhang]

Hi Yuxin,

Your work is really intersting and novel! I want to further understand the detail of your method. However, I have trouble visualizing the watermark pattern. I notice that gt_patch is (1,4,64,64) and I can observe the ring in each channel, while I cannot find that in that init_latents_w, reversed_latents_no_w and reversed_latents_w. Could you help me? Thanks!

[YuxinWenRick]

Hi Qilong, thanks for your interest!

gt_patch contains the ring pattern in every channel as a key, but for init_latents_w, we only inject the key into one of the channels. If you are running python run_tree_ring_watermark.py --run_name no_attack --w_channel 3 --w_pattern ring --start 0 --end 1000 --with_tracking --reference_model ViT-g-14 --reference_model_pretrain laion2b_s12b_b42k, you would expect the pattern is in the last channel since --w_channel 3. Could you double-check if you can see it in the last channel? Also, I don't think we expect any ring pattern for reversed_latents_no_w since it's not watermarked.

Let me know if you have further questions!

[qilong-zhang]

@YuxinWenRick Thanks for your reply. I can obtain the watermark now

[zzzucf]

Hi, I wonder how to visualize the gt_patch, just like what paper shows in Figure 1? It is complex type.

[YuxinWenRick]

Hi @zzzucf , yeah, you are right that we cannot really visualize the complex numbers, so the figure shown in the paper is just the visualization of the real part. You can follow this code to do the visualization.

[zzzucf]

Thank you, that solves my confusion. I really appreciate it.