Setting up γ for ImageNet 256

[Peterande]

Hello, great work, can you give me some advice on setting up γ? I can't summarize the rules well from the table, for example, the difference between FFHQ γ values of different resolutions is too large

[lanqingxi]

1. I have also noticed this problem, in my application, when using FFHQ-256 configured gamma values and strategies, the loss of the discriminator in the pre-training period leads to abnormally large losses, even though lowering that configuration to 15, 1.5, again leads to large losses in the discriminator.
2. my discriminator is based on a pre-trained caformer, is this due to the use of pre-trained weights?

[Blade6570]

Same problem. I have also tried with bigger version of PatchGAN and loss increases abnormally.

Update : Now I don't have the the problem of loss explosion with different gamma values. Most of them works fine. I removed all the normalisation layers (I had one SN by mistake) and borrowed the resblock and weight initialisation trick from R3GAN.