Details about the training of the discriminator

[cuixing100876]

Could you release the training details of the discriminator? The timestep used in diffusion_resampling seems to be in [0,1000], while the timestep used in https://github.com/alsdudrla10/DG/tree/main seems to be in [0,1]

[yujianll]

Thanks for your interests!

The discriminator in our setting is similar to the DG paper, where the input time step is in the range of [0, 1]. For experiments on ImageNet and FFHQ, the noise scale $\sigma$ ranges from 80 to 0, and it is converted to time in [0, 1] at this line. This is the same as the DG paper, and it's based on Equation 163 in the EDM paper.

Please let us know if you have any other questions.