PosterOmni-200K
收藏资源简介:
PosterOmni-200K是由香港科技大学·广州和美团联合构建的大规模多任务海报生成数据集,涵盖局部编辑和全局创作两大范式下的六类任务(如扩展、填充、风格迁移等)。该数据集通过自动化流程生成,包含20万条高质量样本,整合了文本、布局、实体及风格等多模态要素,数据来源包括GPT、Qwen等模型生成的提示词与图像对。其构建过程融合了多模态过滤和任务对齐技术,旨在解决图像到海报生成中语义保真与美学协调的耦合问题,为设计自动化、广告创意等领域提供基准支持。
PosterOmni-200K is a large-scale multi-task poster generation dataset jointly constructed by The Hong Kong University of Science and Technology (Guangzhou) and Meituan. It covers six types of tasks under two paradigms: local editing and global creation, such as expansion, filling, style transfer, etc. Generated through an automated workflow, this dataset contains 200,000 high-quality samples that integrate multimodal elements including text, layout, entities and styles. Its data sources include prompt-image pairs generated by models such as GPT and Qwen. The construction process integrates multimodal filtering and task alignment technologies, aiming to solve the coupled problem of semantic fidelity and aesthetic coordination in image-to-poster generation, and provides benchmark support for fields such as design automation and advertising creativity.




