Zooming into blurry photos to see fine details in 3D

GaussianZoom reconstructs high-fidelity 3D scenes that can be zoomed into 100× magnification, starting from low-resolution input. It uses depth-guided feature alignment to keep views consistent while a vision-language model synthesizes fine details beyond what was originally observed. A new level-of-detail system smoothly transitions between zoom levels without artifacts. Tests on Mip-NeRF360 and Tanks&Temples show it outperforms existing methods in perceptual quality and multi-view consistency under extreme magnification.