DiffFuseNet: Infrared-visible image fusion via diffusion-guided feature alignment and feature consistency alignment

Infrared and visible image fusion aims to synthesize images with thermal radiation and rich texture details. Existing methods suffer from inadequate feature representation and imbalanced fusion due to cross-modal discrepancies, leading to blurred details or insufficient thermal saliency. To address this, we propose DiffFuseNet, which integrates diffusion-guided feature enhancement and explicit feature decoupling. The Dual Diffusion-based Feature Enhancement (D 2 FE) module enhances cross-modal robustness via controllable noise injection and denoising. The Explicit Decoupling and Frequency Decomposition (EDFD) module separates features into shared, modality-specific, and frequency-aware components. Coupled with a frequency-aware fusion mechanism using Haar wavelet and invertible neural networks and a two-stage training strategy, our approach achieves balanced information integration. Experiments on M3FD, RoadScene, and TNO datasets show that DiffFuseNet outperforms state-of-the-art methods in visual quality and objective metrics, with superior detail preservation, thermal saliency, and structural consistency.