Generative-prior super-resolution methods enhance low-resolution inputs using implicit knowledge learned from large-scale models. According to the abstract of the new arXiv paper introducing WorldSR, this approach has a key weakness: those priors are inherently stochastic and tuned for diverse generation, not for faithful grounding in the actual content of the image.
WorldSR is proposed as a way to address that limitation by harnessing world knowledge search for super-resolution. The idea is to move beyond purely generative priors and bring in external, grounded knowledge during the SR process. The available abstract stops mid-sentence, so the paper's specific mechanism and experimental results are not covered in this digest.