COCO-DIMCIM
收藏资源简介:
COCO-DIMCIM数据集是基于COCO数据集构建的,旨在评估文本到图像生成模型的多样性和泛化能力。数据集包含30个对象类别,每个类别有31个种子提示,以及通过大型语言模型生成的属性和属性类型。数据集的创建过程包括从COCO数据集中选择对象类别和提示,然后使用LLM生成属性和属性类型,并使用VQAScore计算属性-概念分数。该数据集适用于评估文本到图像生成模型的多样性和泛化能力。
The COCO-DIMCIM dataset is constructed based on the COCO dataset, designed to evaluate the diversity and generalization capabilities of text-to-image generation models. It encompasses 30 object categories, with 31 seed prompts for each category, along with attributes and attribute types generated by large language models (LLMs). The dataset creation workflow involves selecting object categories and prompts from the COCO dataset, then utilizing LLMs to generate attributes and attribute types, and calculating the attribute-concept score via VQAScore. This dataset is applicable for assessing the diversity and generalization capabilities of text-to-image generation models.
COCO-DIMCIM 数据集概述
数据集简介
- 名称: COCO-DIMCIM Benchmark dataset
- 用途: 用于评估文本到图像生成模型在属性层面的默认模式多样性和泛化能力
- 来源: 基于COCO数据集衍生而来
- 论文: "DIMCIM: A Quantitative Evaluation Framework for Default-mode Diversity and Generalization in Text-to-Image Generative Models"
数据集内容
- 概念数量: 30个
- 属性数量: 494个
- 粗略提示词: 930个
- 密集提示词: 14,641个
文件结构
COCO-DIMCIM/seed_captions_attributes: 包含概念属性和COCO数据集种子描述的json文件COCO-DIMCIM/dense_prompts: 包含从COCO种子描述生成的粗略和密集图像生成提示词的json文件COCO-DIMCIM/simple_attribute_prompts: 包含用于计算属性VQAScores的简单概念-属性提示词的json文件
许可证信息
- 许可证类型: CC-BY-NC
- 使用限制: 仅限作为基准使用
- 附加条款: 使用该数据训练或改进AI模型时,必须在模型名称前加上"Llama"
评估方法
- 使用文本到图像模型生成图像
- 计算DIM分数(默认模式多样性)
- 计算CIM分数(泛化能力)
- 使用提供的Jupyter notebook绘制和分析分数
引用格式
bibtex @misc{teotia2025dimcimquantitativeevaluationframework, title={DIMCIM: A Quantitative Evaluation Framework for Default-mode Diversity and Generalization in Text-to-Image Generative Models}, author={Revant Teotia and Candace Ross and Karen Ullrich and Sumit Chopra and Adriana Romero-Soriano and Melissa Hall and Matthew J. Muckley}, year={2025}, eprint={2506.05108}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2506.05108}, }

- 1DIMCIM: A Quantitative Evaluation Framework for Default-mode Diversity and Generalization in Text-to-Image Generative ModelsMeta AI · 2025年



