In lab, GenR restores degraded faces in 30 seconds

An old family photograph sits in an album, its faces softened by blur, noise and missing pieces. A new system from researchers at the Indian Institute of Technology Gandhinagar and the Indian Institute of Technology BHU, called GenR, produced restored images in about 30 seconds for single-degradation tasks in laboratory evaluations.

The challenge is not simply to sharpen pixels. A face can suffer several forms of damage at once: random brightness or color variations, low resolution, covered sections, blur and compression artifacts created by repeatedly saving an image. Traditional AI restoration often relies on paired training data—one clean image matched with one artificially damaged image—which is expensive to collect for every possible combination of problems.

GenR takes another route. It uses StyleGAN3-based inversion, a process that works backward from a real, damaged image to identify the code for a plausible clean version. The system then optimizes that reconstruction in three stages: first the global structure, identity and pose; then features such as the eyes, nose and jawline; and finally fine textures including skin and hair. Akbar Ali, the IIT Gandhinagar doctoral student who first authored the study, said the controlled move from coarse structure to fine detail is meant to reduce overfitting—the risk of inventing details that do not belong to the original face.

The researchers evaluated GenR on denoising, upsampling, inpainting and deartifacting tasks, covering both single and multiple degradations. They report consistent visual improvements, with single-degradation images produced in about 30 seconds, which they describe as among the fastest results from leading methods. The study appeared in Pattern Recognition Letters.

So what changes in practice? GenR could give archivists a faster way to improve historical photographs and film footage, and could support clearer facial reconstructions in forensic work; it could also improve images on video-conferencing and social-media platforms. But it remains a laboratory framework, not an identity machine: when the source image is severely damaged, it can create a realistic face that does not match the original person, while complex settings can produce unrealistic results without strict regularization.