I have an image autoencoder model trained as follows:

Step 1) train a GAN to obtain a generator capable of drawing from the data manifold by sampling a normal distribution in latent space

Step 2) train an encoder on the front, keeping the generator frozen, effectively mapping images into the normally distributed latent space.

Is there a name/reference for such a model?

I know it's not really a VAE-GAN because it has no variational component, and the decoder and encoder are trained separately. It's not quite a vanilla AE either since the latent space is structured and constrained in the GAN training step. The models I get from this approach, while not SoTA, do have desirable qualities like disentangled representations but without the instabilities of VAE-GAN training.

I do use a perceptual loss when training the encoder, using intermediate feature maps from the GAN-trained discriminator, but I don't think it makes a big difference to the question I'm asking.



You must log in to answer this question.