T2I-ConBench
收藏资源简介:
T2I-ConBench是一个用于持续微调文本到图像模型的基准数据集,由上海交通大学和华为公司的研究人员创建。该数据集包含约100,000条数据,涵盖了个性化对象生成和特定领域图像质量提升两种场景。数据集通过合成数据生成和人工筛选的方式获取,并包含对模型持续学习、遗忘和跨任务泛化能力的评估指标。该数据集旨在解决文本到图像模型在持续微调过程中可能出现的知识遗忘和跨任务泛化问题。
T2I-ConBench is a benchmark dataset for continual fine-tuning of text-to-image models, developed by researchers from Shanghai Jiao Tong University and Huawei. This dataset contains approximately 100,000 samples, covering two scenarios: personalized object generation and image quality enhancement in specific domains. It is collected through synthetic data generation and manual filtering, and includes evaluation metrics for measuring models' continual learning, catastrophic forgetting, and cross-task generalization capabilities. This dataset aims to address the issues of knowledge forgetting and cross-task generalization that may arise during the continual fine-tuning of text-to-image models.
DreamBooth数据集概述
数据集基本信息
- 名称: DreamBooth Dataset
- 用途: 用于主题驱动的文本到图像扩散模型的微调
- 来源: Google论文《DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation》的官方数据集
数据集内容
- 主题数量: 30个(15个不同类别)
- 活体主题: 9个(狗和猫)
- 物体主题: 21个
- 图像数量: 每个主题4-6张图像
- 图像特点: 在不同条件、环境和角度下拍摄
文件说明
- prompts_and_classes.txt: 包含论文中用于活体主题和物体的所有提示词及类别名称
- references_and_licenses.txt: 包含所有来自www.unsplash.com的图像的参考链接、摄影师署名和图像许可证信息
数据来源
- 图像由论文作者拍摄或来自www.unsplash.com
学术引用
bibtex @inproceedings{ruiz2023dreambooth, title={Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation}, author={Ruiz, Nataniel and Li, Yuanzhen and Jampani, Varun and Pritch, Yael and Rubinstein, Michael and Aberman, Kfir}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, year={2023} }
免责声明
- 非Google官方支持产品

- 1T2I-ConBench: Text-to-Image Benchmark for Continual Post-training上海交通大学 · 2025年



