DEIG-Bench
收藏资源简介:
DEIG-Bench是由中山大学、复旦大学和耶鲁大学联合构建的精细化语义控制多实例生成基准数据集。该数据集针对现有数据集中人类实例不足和单属性提示依赖的缺陷,提供了包含颜色、材质、纹理等多属性组合的描述标注,并通过视觉语言模型生成高质量实例级描述。数据集特别设计了人体穿戴区域的颜色组合性评估和物体属性复杂度递增的测试场景,采用双VLM问答机制验证语义一致性,旨在解决复杂文本描述下的细粒度视觉生成问题。
DEIG-Bench is a fine-grained semantic-controlled multi-instance generation benchmark dataset co-developed by Sun Yat-sen University, Fudan University and Yale University. Aiming at the drawbacks of insufficient human instances and single-attribute prompt dependence in existing datasets, this dataset provides descriptive annotations with multi-attribute combinations including color, material, texture and others, and generates high-quality instance-level descriptions via vision-language models (VLMs). The dataset specially designs test scenarios covering color combinability evaluation of human wearing areas and gradually increasing complexity of object attributes, and adopts a dual-VLM question-answering mechanism to verify semantic consistency, aiming to solve the fine-grained visual generation problem under complex text descriptions.
- 1DEIG: Detail-Enhanced Instance Generation with Fine-Grained Semantic Control中山大学; 复旦大学; 耶鲁大学 · 2026年



