Lightweight image super-resolution with tokenized dynamic embedding network

Abstract

Image super-resolution is a crucial task in computer vision, aiming to reconstruct high-resolution images from low-resolution counterparts. Despite the remarkable progress of deep learning-based methods, existing approaches often face challenges in balancing reconstruction quality, computational efficiency, and model compactness. In this paper, we propose a novel tokenized dynamic embedding network, which integrates adaptive feature tokenization and dynamic embedding mechanisms to enhance super-resolution performance while maintaining efficiency. Specifically, we employ an adaptive feature tokenization strategy to selectively extract essential tokens, reducing computational complexity while preserving key image details. Additionally, we introduce a dynamic context embedding attention module for efficient long-range dependency modeling and a dual-perspective feature integration module for integrating spatial and contextual information, ensuring both fine-grained textures and global consistency. Extensive experiments on benchmark datasets demonstrate that our method outperforms state-of-the-art lightweight models in terms of objective metrics and perceptual quality, while maintaining a compact and efficient design suitable for real-world applications. The source code is available at https://github.com/zxycs/TDEN.