Disentangled Inference for GANs With Latently Invertible Autoencoder

Generative Adversarial Networks (GANs) can synthesize more and realistic images. However, one fundamental issue hinders their practical applications: the incapability of encoding real samples in latent space. Many semantic image editing applications rely on inverting given into space then manipulating inverted code. One possible solution is to learn an encoder for GAN via Variational Auto-Encoder. entanglement poses a major challenge learning encoder. To tackle enable inference GANs, we propose novel method named Latently Invertible Autoencoder (LIA). In LIA, invertible network its inverse mapping are symmetrically embedded autoencoder. The decoder LIA first trained as standard with network, learned from disentangled autoencoder by detaching LIA. It thus avoids problem caused Extensive experiments FFHQ face dataset three LSUN datasets validate effectiveness inversion applications. Code models available at https://github.com/genforce/lia .