IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis

We present a novel introspective variational autoencoder (IntroVAE) model for high-resolution photographic images. IntroVAE is capable of-evaluating the quality of its generated samples and improving itself. Its inference and generator models are jointly trained in an way. On one hand, the generator is required to reconstruct the images from the noisy outputs of the inference model as normal VAEs. On other hand, the inference model is encouraged to classify between the and real samples while the generator tries to fool it as GANs. These famous generative frameworks are integrated in a simple yet efficient-stream architecture that can be trained in a single stage. IntroVAE the advantages of VAEs, such as stable training and nice latent. Unlike most other hybrid models of VAEs and GANs, IntroVAE requires extra discriminators, because the inference model itself serves as a to distinguish between the generated and real samples. demonstrate that our method produces high-resolution-realistic images (e. g. , CELEBA images at \\ (1024^2\\) ), which are to or better than the state-of-the-art GANs.