Omni-Edit Benchmark

该数据集是一个手动创建的基准，包含了来自Pexels和LAION-5B的62张图片，旨在评估基于指令的图像编辑在多个任务和不同宽高比下的表现。所选取的图片在美学评分上较高，且场景多样化，覆盖了各种编辑任务。规模上，该数据集包含了62张图片和434项编辑工作，其任务是评估基于指令的图像编辑效果。This manually constructed benchmark dataset contains 62 images sourced from Pexels and LAION-5B, designed to evaluate instruction-based image editing across multiple tasks and varying aspect ratios. The selected images feature high aesthetic scores and cover diversified scenarios, suitable for a wide array of editing tasks. In terms of scale, the dataset comprises 62 images and 434 editing tasks, with the core goal of assessing the effectiveness of instruction-based image editing.

搜集汇总

背景与挑战

背景概述

OMNI-EDIT-Filtered-1.2M是一个大规模图像编辑数据集，包含120万对图像编辑对，覆盖添加、替换、移除等七种编辑技能。该数据集通过专家模型蒸馏和大型多模态模型过滤合成，具有高多样性和高质量，支持任意分辨率和宽高比，并经过VIEScore过滤以确保数据质量。