Underwater image enhancement is fundamental to ocean exploration, as it can reduce the distortion caused by refraction and absorption in underwater imaging. Although many deep-learning-based underwater image enhancement methods achieved good performance based on multiple paired datasets, these methods can easily suffer from domain shifts and thus exhibit poor generalization to unfamiliar waters. In this work, we report a novel unpaired domain-irrelevant underwater image enhancement framework termed Unpaired Domain-Irrelevant Underwater Image Enhancement (UDUIE), with a domain-irrelevant loss built upon contrastive learning. The framework enhances domain generalization capability by transforming sample features and separating content features from water features in degraded photographs. Furthermore, we also propose a real-world, large-scale underwater image dataset (RLUI) in which degraded and clear domain images originate from existing photos, containing 7127 degraded images and 5644 clear images. Extensive experiments demonstrate that our approach outperforms current state-of-the-art methods in underwater image enhancement tasks across different underwater datasets.