Human-Aligned MLLM Judges for Fine-Grained Image Editing Evaluation Benchmark
收藏资源简介:
该数据集由多所顶尖高校和研究机构联合构建,旨在通过细粒度多模态大语言模型(MLLM)评估框架解决图像编辑模型评估的挑战。数据集整合了人类评估、MLLM评分、模型输出及传统指标,覆盖图像保留、编辑质量和指令忠实度等12个维度。其创新性在于将主观评价分解为可解释的语义因子,并通过大规模人类研究验证了MLLM评估与人类判断的高度一致性。该资源支持离线(有真值)和在线(无真值)双场景,为图像编辑算法的研究、比较与优化提供了标准化基准。
This dataset was jointly constructed by multiple top-tier universities and research institutions, aiming to address the challenges in evaluating image editing models via a fine-grained multimodal large language model (MLLM)-based evaluation framework. The dataset integrates human evaluations, MLLM scores, model outputs and traditional metrics, covering 12 dimensions including image preservation, editing quality, instruction faithfulness and others. Its innovation lies in decomposing subjective evaluations into interpretable semantic factors, and verifying through large-scale human studies that MLLM-based evaluations have a high degree of consistency with human judgments. This resource supports two scenarios: offline (with ground truth) and online (without ground truth), providing a standardized benchmark for the research, comparison and optimization of image editing algorithms.
- 1Human-Aligned MLLM Judges for Fine-Grained Image Editing Evaluation: A Benchmark, Framework, and Analysis弗吉尼亚大学; 哥伦比亚大学; 范德比尔特大学; 奥多比研究院; 杜比实验室; 思科研究院; 南加州大学; 威斯康星大学麦迪逊分校; 俄勒冈大学; 德州农工大学 · 2026年



