FiVA
收藏资源简介:
FiVA数据集是由斯坦福大学等机构创建的细粒度视觉属性数据集,旨在为文本到图像扩散模型提供高质量的视觉属性标注。该数据集包含约100万张高分辨率生成图像,每张图像都标注了多种视觉属性,如颜色、光照、纹理等。数据集的创建过程包括属性定义、提示生成、LLM过滤和人工验证,确保了数据的高质量和多样性。FiVA数据集的应用领域广泛,主要用于提升图像生成模型的可控性和用户定制能力,解决现有模型在细粒度视觉属性控制上的不足。
The FiVA dataset is a fine-grained visual attribute dataset created by Stanford University and other institutions, aiming to provide high-quality visual attribute annotations for text-to-image diffusion models. It contains approximately 1 million high-resolution generated images, each annotated with multiple visual attributes such as color, lighting, texture and so on. The dataset construction workflow includes attribute definition, prompt generation, LLM-based filtering and manual verification, which ensures the high quality and diversity of the dataset. The FiVA dataset has a wide range of application scenarios, and is mainly used to improve the controllability and user customization capabilities of image generation models, addressing the shortcomings of existing models in fine-grained visual attribute control.

- 1FiVA: Fine-grained Visual Attribute Dataset for Text-to-Image Diffusion Models斯坦福大学, 香港中文大学, 浙江大学, S-Lab, NTU, 上海人工智能实验室, CPII under InnoHK · 2024年



