YinYangAlign
收藏资源简介:
YinYangAlign是由南卡罗来纳大学人工智能研究所和Meta AI共同创建的高级基准测试框架,旨在系统量化文本到图像系统(T2I)的对齐保真度。该数据集包括详细的对齐公理数据集,其中包含人类提示、选定响应、被拒绝的AI生成输出以及对 underlying矛盾的说明。该数据集涉及六个基本且固有的矛盾设计目标,每个目标都代表了图像生成中的关键张力。通过结合先进的视觉语言模型和人类验证,YinYangAlign为评估和优化T2I系统提供了一种全面的方法。
YinYangAlign is an advanced benchmarking framework co-created by the Artificial Intelligence Institute of the University of South Carolina and Meta AI, designed to systematically quantify the alignment fidelity of text-to-image (T2I) systems. This dataset includes a detailed alignment axiom dataset containing human prompts, selected responses, rejected AI-generated outputs, and explanations for the underlying contradictions. It covers six fundamental and inherently contradictory design objectives, each representing a key tension in image generation. By combining advanced vision-language models and human validation, YinYangAlign provides a comprehensive approach for evaluating and optimizing T2I systems.

- 1YINYANG-ALIGN: Benchmarking Contradictory Objectives and Proposing Multi-Objective Optimization based DPO for Text-to-Image Alignment南卡罗来纳大学人工智能研究所 · 2025年



