Low bit rate generative face compression using inverse GAN

The paper investigates the adoption of generative adversarial networks (GANs) for low-bit-rate learned image compression, aiming to both preserve fine visual details and maintain semantic and biometric consistency. The designed strategy is based on the inversion of a StyleGAN image generation network and the characterisation of noise and visual features through strong quantisation and a side information channel. Experimental results show that the images can be stored with a lower amount of bits and present better visual details with respect to standard state-of-the-art learned coding schemes.