How to keep fuzzy photos faithful while adding missing detail

Real-world photo upscaling faces a tension: stay true to the blurry original or synthesize realistic details that might not exist. VICR, a diffusion transformer model, splits this problem into two decoupled paths—one preserving structure, one generating realistic detail—with an adaptive refinement step for severely degraded regions. Tests across standard benchmarks show it outperforms existing methods while using fewer parameters.