Inverting Adversarially Robust Networks for Image Synthesis

Despite unconditional feature inversion being the foundation of many image synthesis applications, training an inverter demands a high computational budget, large decoding capacity and imposing conditions such as autoregressive priors. To address these limitations, we propose use adversarially robust representations perceptual primitive for inversion. We train encoder to extract disentangled perceptually-aligned representations, making them easily invertible. By simple generator with mirror architecture encoder, achieve superior reconstruction quality generalization over standard models. Based on this, autoencoder demonstrate its improved performance style transfer, denoising anomaly detection tasks. Compared recent ImageNet methods, our model attains significantly less complexity. Code available at https://github.com/renanrojasg/adv_robust_autoencoder .