How this field actually works and where it is going, written by the people building on it, citing the primary papers. No listicles.
Restoration is an inverse problem with no unique answer. From one-network-per-defect GANs to diffusion posterior sampling to instruction-driven latent models, and why every output is plausible rather than true. 20 primary sources.
29 July 2026
A retoucher's afternoon in 2015; cents of GPU time in 2026. How general image priors ate the specialist stack, how hosted inference ate the infrastructure, and what stubbornly refused to become a commodity.
29 July 2026
The first restoration is free, straight from your browser.
