Generative models with kernel distance in data space

Generative models dealing with modeling a joint data distribution are generally autoencoder, or GAN based. Both have pros and cons, generating blurry images being unstable in training. We propose new generative model resembling classical transforming Gaussian noise into space without adversarial optimization. Training of the proposed is two-step procedure. First, we train an autoencoder-based architecture to manifold. Second, use Latent Trick map autoencoder’s latent space. The resulting Cramer-Wold (LCW) generator achieves competitive scores. Elimination training replacing discriminator kernel methods results stable procedure that not prone mode collapse. also show introduced can improve capabilities other latent-based models. validate on standard benchmarks compare it different approaches.