MagicBrush

该数据集是首个大规模的手动标注数据集，专门用于指导真实图像编辑，覆盖了多种场景，包括单轮对话、多轮对话、提供遮罩和不提供遮罩的编辑。该数据集包含了超过10,000个手动标注的三元组（源图像、指令和目标图像）。此外，它支持训练大规模的文本指导图像编辑模型，并已用于广泛评估当前的图像编辑基线。规模上，该数据集拥有超过10,000个三元组，其任务是指导式图像编辑。

This is the first large-scale manually annotated dataset specifically designed for instructional real-world image editing. It covers a wide range of scenarios, including single-turn dialogue, multi-turn dialogue, image editing with and without masks. The dataset contains over 10,000 manually annotated triplets (source image, editing instruction, and target image). Additionally, it enables the training of large-scale text-guided image editing models and has been used to extensively evaluate existing image editing baselines. In terms of scale, this dataset includes over 10,000 triplets focused on instructional image editing tasks.