Do giant image generators actually fix blurry photos?

Large generative models excel at creating images from scratch but struggle with pixel-level precision tasks like deblurring and denoising. LL-Bench evaluates 10 state-of-the-art generative models and 21 conventional restoration tools on 16 real degradation types using 152,020 expert preference annotations. The benchmark reveals generative models frequently add fake details rather than recover lost information—a failure mode conventional methods avoid. The authors also propose LL-Score, an evaluator that better aligns with human judgments than standard metrics.