CompBench
收藏资源简介:
CompBench是一个大规模的基准数据集,专门为复杂指令引导的图像编辑任务而设计。该数据集包含了3,000多个图像-指令对,涵盖了九种不同的图像编辑任务类别,如对象添加、对象删除、对象替换、多对象编辑、多轮编辑、隐式推理、动作编辑、位置编辑和视角编辑。数据集的内容源自真实世界场景,具有高度复杂性和多样性,能够全面评估图像编辑模型在精确操作能力方面的表现。数据集的创建采用了多轮专家审查,确保了编辑质量,并使用了一种指令解耦策略,将编辑意图分解为四个关键维度:位置、外观、动态和对象,以确保指令与复杂编辑需求之间的紧密对齐。CompBench旨在解决现有图像编辑模型在评估复杂场景理解和细微操作能力方面的局限性,并为下一代指令引导的图像编辑系统的开发提供关键见解。
CompBench is a large-scale benchmark dataset specifically designed for complex instruction-guided image editing tasks. This dataset contains over 3,000 image-instruction pairs, covering nine distinct image editing task categories, including object addition, object removal, object replacement, multi-object editing, multi-turn editing, implicit reasoning, action editing, position editing, and viewpoint editing. The dataset content originates from real-world scenarios, boasts high complexity and diversity, and can comprehensively evaluate the performance of image editing models in terms of precise manipulation capabilities. The dataset construction adopted multi-round expert reviews to ensure editing quality, and employed an instruction decoupling strategy that decomposes editing intent into four critical dimensions: position, appearance, dynamics, and object, to ensure tight alignment between instructions and complex editing requirements. CompBench aims to address the limitations of existing image editing models in evaluating complex scene understanding and fine-grained manipulation capabilities, and provide key insights for the development of next-generation instruction-guided image editing systems.

- 1CompBench: Benchmarking Complex Instruction-guided Image Editing华东师范大学 · 2025年



