Hello authors,
Could you please let me know what EDM arch and other network config you used for training watermarked ImageNet images? I am training on a subset of imagenet (69 classes, 64 x 64 res) and both arch=ddpmpp and adm are generating images where the watermark decoder is not able to detect any watermarks. Also, loss on ADM after 25000 kimg probably became nan since the generated images are all black.
I also tested the code for FFHQ dataset on which it is working perfectly fine -- perfect generations as well as perfect watermark detection. arch was ddpmpp as mentioned in your readme. But for imagnet, I am running into the above-mentioned troubles. Please help me.
Best,
Avideep
Hello authors,
Could you please let me know what EDM arch and other network config you used for training watermarked ImageNet images? I am training on a subset of imagenet (69 classes, 64 x 64 res) and both arch=ddpmpp and adm are generating images where the watermark decoder is not able to detect any watermarks. Also, loss on ADM after 25000 kimg probably became nan since the generated images are all black.
I also tested the code for FFHQ dataset on which it is working perfectly fine -- perfect generations as well as perfect watermark detection. arch was ddpmpp as mentioned in your readme. But for imagnet, I am running into the above-mentioned troubles. Please help me.
Best,
Avideep