Question regarding GAN-only

[kunwooch]

Hi Alex,

I've recently looked into GAN models, and I was wondering where the GAN-only model you used in the paper is based on. The paper mentions that it is representative of prior stochastic GAN-based methods, and it seems to me that the overall structure resembles pix2pix (as they are both conditional GAN). Is the GAN-only model same as the SAVP without VAE?

Thank you.

[alexlee-gk]

The GAN-only model is not based exactly on any single prior model. The generator is similar to what has previously been used for video prediction (not pix2pix). The discriminator is based on the one from SNGAN except that the conv2d layers have been replaced with conv3d layers in order to discriminate videos instead of single images.

Yes, the GAN-only model is the same as SAVP but without the VAE (and without the second discriminator that discriminate videos sampled from the posterior).