Visualignment/CoProv2-SDXL
收藏官方服务：
资源简介：
CoProV2是一个合成生成的有害和安全图像-文本对的数据集，旨在用于文本到图像生成模型中的安全性目的的直接偏好优化（DPO）。它通过训练“安全专家”来引导生成过程远离特定的安全相关概念，实现可扩展的安全性对齐。
CoProV2 is a synthetically generated dataset of harmful and safe image-text pairs, designed for Direct Preference Optimization (DPO) for safety purposes in Text-to-Image (T2I) models. It facilitates the training of safety experts to guide the generative process away from specific safety-related concepts, enabling scalable safety alignment.
提供机构：
Visualignment
