creative-graphic-design/AesEvalBench
收藏资源简介:
AesEval-Bench是一个用于评估视觉语言模型是否能评估平面设计美学的基准测试。该基准测试包含扰动后的平面设计样本,包括预览图像、高亮预览、元素级元数据和美学问题标签。基准测试涵盖四个维度和十二个指标:布局(平衡、层次、留白、对齐)、排版(可读性、层次)、图形(质量、相关性)和颜色(和谐、对比、吸引力、心理学)。原始评估代码支持二元美学判断、问题区域选择和边界框定位,而Hugging Face加载器在单个test分割中公开基准样本和标签。
AesEval-Bench is a benchmark for evaluating whether vision-language models can assess graphic design aesthetics. The benchmark contains perturbed graphic design samples with preview images, highlighted previews, element-level metadata, and labels for aesthetic issues. The benchmark spans four dimensions and twelve indicators: Layout (balance, layering, whitespace, alignment), Typography (legibility, hierarchy), Graphics (quality, relevance), and Color (harmony, contrast, appeal, psychology). The original evaluation code supports binary aesthetic judgment, problematic-region choice, and bounding-box localization. This Hugging Face loader exposes the benchmark samples and labels in a single `test` split.




