Reason50K
收藏资源简介:
Reason50K是一个大规模数据集,专门用于训练和评估基于假设性指令推理的图像编辑。该数据集包含超过5.1万个样本,涵盖物理、时间、因果和故事推理四个关键推理场景。每个样本由一个源图像、一个假设性指令和一个反映预期编辑的目标图像组成。数据集采用逆向策略构建,即从目标图像生成源图像,并利用GPT生成假设性指令。Reason50K旨在支持基于假设性指令的推理,使图像编辑模型能够理解并执行复杂的编辑任务。
Reason50K is a large-scale dataset specifically developed for training and evaluating image editing models that conduct reasoning based on hypothetical instructions. This dataset includes over 51,000 samples covering four critical reasoning scenarios: physical, temporal, causal, and narrative reasoning. Each sample comprises a source image, a hypothetical instruction, and a target image that represents the expected editing outcome. The dataset is built using a reverse strategy, where source images are generated from target images, and hypothetical instructions are produced with GPT. Reason50K is intended to support reasoning based on hypothetical instructions, allowing image editing models to comprehend and perform complex editing tasks.
数据集概述
基本信息
- 数据集名称: Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning
- 相关论文: Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning
- 作者: Qingdong He, Xueqin Chen, Chaoyi Wang, Yanjie Pan, Xiaobin Hu, Zhenye Gan, Yabiao Wang, Chengjie Wang, Xiangtai Li, Jiangning Zhang
- 机构: Youtu Lab (Tencent), TU Delft, University of Chinese Academy of Sciences, Fudan University, Nanyang Technological University
数据集状态
- 发布计划: 代码、模型和数据集将于2025年9月前在Huggingface上发布。
引用信息
- BibTeX: bibtex @article{he2025reasoning, title={Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning}, author={He, Qingdong and Chen, Xueqin and Wang, Chaoyi and Pan, Yanjie and Hu, Xiaobin and Gan, Zhenye and Wang, Yabiao and Wang, Chengjie and Li, Xiangtai and Zhang, Jiangning}, journal={arXiv preprint arXiv:2507.01908}, year={2025} }
联系方式
- 联系人: Qingdong He
- 邮箱: yingcaihe@tencent.com

- 1Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning腾讯优图实验室 · 2025年



