BiComp
收藏资源简介:
BiComp是由复旦大学研究团队构建的大规模高质量文本-图像偏好数据集,专门用于增强组合式文本到图像生成模型的性能。该数据集包含57,474张原始图像和94,502张编辑后图像,总计151,976个样本,覆盖颜色、形状、纹理、空间关系、非空间关系和计数六大组合维度,数据来源于CONPAIR、T2I-CompBench等多个公开语料库。数据集通过自动化流水线构建,包括提示收集、图像生成、描述重生成、编辑和VQA过滤等严格质量控制步骤。该数据集主要应用于组合式文本到图像生成领域,旨在解决现有模型在属性绑定、对象关系和计数等复杂组合提示下生成准确图像的难题。
BiComp is a large-scale high-quality text-image preference dataset constructed by a research team from Fudan University, specifically designed to enhance the performance of compositional text-to-image generation models. This dataset contains 57,474 raw images and 94,502 edited images, totaling 151,976 samples, covering six compositional dimensions: color, shape, texture, spatial relations, non-spatial relations, and counting. The data is sourced from multiple public corpora such as CONPAIR and T2I-CompBench. The dataset is built through an automated pipeline with strict quality control steps including prompt collection, image generation, description regeneration, editing and VQA filtering. It is primarily applied in the field of compositional text-to-image generation, aiming to address the challenge that existing models face when generating accurate images under complex compositional prompts such as attribute binding, object relations and counting.
数据集概述:BiComp
名称:BiComp
来源机构:复旦大学(上海智能信息处理重点实验室、上海智能视觉计算协同创新中心)
发布形式:通过 Hugging Face 发布,数据集地址为 https://huggingface.co/datasets/anzeameol/BiComp
发布时间:2026年3月23日
数据集用途
BiComp 是面向 组合式文本到图像生成(Compositional Text-to-Image Generation)任务的数据集,用于训练和评估 BiDPO(区域感知双模态直接偏好优化)方法,该工作已被 CVPR 2026 接收。
数据集内容
- 用于支持 区域感知的双模态直接偏好优化 训练过程,提升模型对文本提示中多个物体、属性及空间关系等组合语义的生成能力。
- 数据集包含图像-文本对,并配合 VisMin 数据集(需单独下载)共同使用。
数据集下载方式
-
使用 Hugging Face 命令行工具下载: bash hf download "anzeameol/BiComp" --repo-type "dataset" --local-dir "./datasets/BiComp"
-
或运行项目提供的下载脚本: bash bash ./scripts/download/download_BiComp.sh
相关资源与引用
-
项目代码仓库:https://github.com/anzeameol/BiDPO
-
预训练检查点(SDXL-BiDPO):https://huggingface.co/anzeameol/sdxl-BiDPO
-
参考文献(BiDPO 论文,CVPR 2026):
@inproceedings{liucompositional, title={Compositional Text-to-Image Generation Via Region-aware Bimodal Direct Preference Optimization}, author={Liu, Zhuohan and Peng, Wujian and Chen, Yitong and Wu, Zuxuan}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, year={2026} }




