From what I understand in GAN, the main idea is that you have a generator and a discriminator network that are "competing" with each other. The generator trying to make images that the discriminator is not able to distinguish from real images. Since recently diffusion models started to become more popular. I haven't seen many papers showing the DDPM being the generator in a GAN system. Is there a reason for this, since from what I understand pretty much any model could be the generator?
Asked
Active
Viewed 109 times
1
nbro
- 39,006
- 12
- 98
- 176
Nikita Belooussov
- 123
- 3
-
What would, in your view, be the use of having a (trained?) DDPM generate images to be then discriminated by the GAN? Im not following what you would want to accomplish. Do you want to train the DDPM using the adversarial loss of the GAN training procedure? – Robin van Hoorn Feb 23 '23 at 09:46