EAPFusion: Intrinsic evolving auxiliary prior guidance for infrared and visible image fusion

Infrared-visible image fusion aims to synthesize thermal saliency and visible textures into an informative representation for robust real-world perception. Existing methods often rely on static learned weights with limited test-time adaptivity and suffer from granularity mismatch when coarse auxiliary semantics are injected into low-level fusion features, hindering simultaneous target enhancement and detail preservation. To address these limitations, we propose EAPFusion, a self-evolving intrinsic-prior-guided framework that dispenses with external auxiliary models. EAPFusion maintains compact intrinsic priors and progressively updates them across scales to provide granularity-aligned guidance. The evolved priors are further transformed into instance-adaptive convolutional kernels through prior-driven dynamic convolution, enabling content-aware local modulation beyond fixed filters. In addition, a shuffle-guided channel fusion module interleaves infrared and visible channels and performs local channel mixing to enhance cross-modal complementarity. Extensive experiments on multiple datasets, including cross-dataset evaluation and downstream semantic segmentation, demonstrate that EAPFusion achieves superior quantitative and qualitative fusion performance while improving downstream semantic understanding. The source code for this work is publicly available at: https://github.com/Zhenyu-Sun-86587/EAPFusion.