4KAgent

Why we included this project

Most upscalers apply one fixed model to every image, which falls apart when a collection mixes old photos, heavy compression, and real-world noise. 4KAgent takes a different route: a vision-language agent inspects the input, figures out what kind of degradation it is dealing with, and picks a restoration path before upscaling to 4K. That planning, execution, and rollback loop is what lets it handle distorted or heavily compressed sources that a one-size-fits-all tool would mangle, so it suits archivists and photographers working through mixed or legacy image sets. The repo includes inference code, example profiles for classic, real-world, old-photo, and multi-degradation restoration, and the DIV4K-50 evaluation dataset, so you can reproduce the NeurIPS 2025 results and try it on your own images before adopting it.