anonymousllbench/llbench-dataset

LL-Bench是一个大规模人类偏好基准，旨在评估大规模生成模型时代的低层视觉恢复。它比较了10个大规模生成模型、16个专业模型和5个全能模型，覆盖16个低层视觉任务，包括运动去模糊、阴影去除、去雪、去雨、超分辨率、HDR成像、低光增强、老照片修复、压缩伪影去除、雨滴去除、水下增强、去雾、去噪、散焦去模糊、光晕去除和反射去除。数据集包含密集的人类标注，如成对质量偏好、Bradley–Terry分数和每幅图像的幻觉标签，用于全面评估模型性能。数据集结构包括元数据文件、人类偏好标注、源图像和恢复图像，适用于图像处理和视觉研究。LL-Bench is a large-scale human-preference benchmark for evaluating low-level vision restoration in the era of large generative models. It compares 10 large generative models, 16 specialist models, and 5 all-in-one models across 16 low-level vision tasks, including motion deblurring, shadow removal, desnow, derain, super-resolution, HDR imaging, low-light enhancement, old-photo restoration, compression artifact removal, raindrop removal, underwater enhancement, dehaze, denoise, defocus deblurring, flare removal, and reflection removal. The dataset includes dense human annotations such as pairwise quality preferences, Bradley–Terry scores, and per-image hallucination labels, providing a comprehensive evaluation framework. It is structured with metadata, human preference data, source images, and restored images for research in image processing and computer vision.