LL Bench: Rethinking Low Level Vision Evaluation in the Era of Large Scale Generative Models via Human Preferences Anonymous release prepared for NeurIPS 2026 review. Please do not redistribute. LL Bench is a large scale, human preference benchmark for evaluating low level vision restoration in the era of large generative models (LGMs). It compares 10 LGMs with 16 specilist and 5 all in one models across 16 low level vision tasks , paired with dense human annotations:pairwise quality preferences, Bradley–Terry scores, and per image hallucination labels. Tasks Motion Deblurring, Shadow Removal, Desnow, Derain, Super Resolution, HDR Imaging, Low Light Enhancement, Uncompleted (old photo restoration), Compression Artifact Removal, Raindrop Removal, Underwater Enhancement, Dehaze, Denoise, Defocus Deblurring, Flare Removal, Reflection Removal. Directory layout Usage Quick load Resolving an image Joining a pair to its images Notes Annotators were instructed to flag over processing / hallucination separately from preference ranking; the two label types are independent. wo gt trials have empty gt image path ; check has gt before reading. Some task × dataset combinations contain only a sub…
Runs entirely in your browser via DuckDB-Wasm — this dataset's real data file is loaded once, then queried locally. Nothing is sent to a server.
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy