AesGuide
收藏资源简介:
AesGuide是由北京大学团队构建的首个面向美学指导任务的大规模数据集,包含10,748张真实场景照片,每张均标注了美学评分、专业分析和拍摄改进建议。该数据集通过爬取网络平台照片与专业摄影师合作采购双重渠道构建,采用两阶段标注框架(MLLM初步提炼+专家人工修正)确保质量,重点解决现有模型在美学缺陷识别和可操作建议生成方面的不足。其核心应用领域为计算美学,旨在通过数据驱动方式提升多模态大模型在拍摄指导(如构图优化、光线调整)和后期裁剪中的解释性与交互能力。
AesGuide is the first large-scale dataset dedicated to aesthetic guidance tasks, constructed by a team from Peking University. It comprises 10,748 real-world scene photographs, each annotated with aesthetic scores, professional analyses, and actionable shooting improvement suggestions. This dataset is built through two channels: crawling photos from online platforms and procuring relevant materials in collaboration with professional photographers. It adopts a two-stage annotation framework (MLLM-based preliminary extraction + expert manual revision) to ensure data quality, primarily addressing the limitations of existing models in aesthetic defect recognition and actionable suggestion generation. Its core application domain is computational aesthetics, aiming to enhance the interpretability and interactivity of multimodal large language models in shooting guidance such as composition optimization, lighting adjustment and post-production cropping through data-driven approaches.
- 1Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping北京大学·计算机技术研究所 · 2026年



