您当前的位置: > 详细浏览

CPA-IQA:A Contextual Preference-Aligned Framework for AIGC Quality Assessment

请选择邀稿期刊:
摘要: With the rapid advancement of AI-Generated Content (AIGC) technologies, evaluating the quality of generated images has become increasingly critical, particularly in specialized domains such as cultural heritage preservation. This paper presents CPA-IQA (Contextual Preference-Aligned Image Quality Assessment), a novel no-reference multi-modal evaluation framework that addresses the limitations of existing quality assessment methods. Unlike traditional approaches that focus solely on low-level perceptual quality or semantic alignment, CPA-IQA integrates hierarchical multi-modal alignment, distortion identification, and scenario-adaptive preference modeling to provide comprehensive quality assessment tailored to specific application contexts. Our experimental validation on a dataset of 192 heritage craft images generated by three state-of-the-art models (Doubao, Kwan 2.6, and JiMengAI 3.0) demonstrates the framework’s effectiveness in distinguishing quality differences across diverse scenarios and identifying AIGC-specific artifacts. The results show that Kwan 2.6 achieves the highest average quality score (43.75), followed by JiMengAI 3.0 (43.37) and Doubao (41.60), with significant variations across different cultural themes and artistic styles.

版本历史

[V2] 2026-10-19 21:55:15 ChinaXiv:202602.00217V2 下载全文
[V1] 2026-02-05 22:48:04 ChinaXiv:202602.00217v1 查看此版本 下载全文
点击下载全文
预览
同行评议状态
待评议
许可声明
metrics指标
  •  点击量4115
  •  下载量1071
评论
分享
申请专家评阅