GenR: Generative latent inversion for blind face restoration

Akbar Ali, Indra Deep Mastan, Shanmuganathan Raman

Abstract: Blind face restoration aims to recover high‑quality images from inputs corrupted by noise, compression artifacts, missing content, and low resolution. Paired supervision is often unavailable, and high-quality restoration with an unsupervised alternative is challenging under the composite degradations. We introduce GenR, an unsupervised framework that leverages pretrained StyleGAN3 priors and optimizes a latent code so that the degraded rendering of the synthesized image matches the observation. Our design combines (i) alias‑free SG3 synthesis for stable, geometry‑consistent inversion;(ii) staged latent expansion W→ W+→ W++ with frequency‑aware regularization to balance capacity and overfitting; and (iii) a masked multiscale perceptual–structural loss that aggregates symmetric LPIPS across pyramid levels, l 1, gradient, and ssim. We evaluate denoising, super-resolution, inpainting, deartifacting, and multi-degradation chains across severity levels (XS–XL) using standard non-reference metrics. GenR delivers competitive or superior performance to relevant methods and produces high-quality restoration for single and composite degradations.