SVSFusion: Semantic-prior visual self-prompted learning for infrared and visible image fusion

Infrared and visible image fusion (IVIF) aims to integrate complementary information from heterogeneous modalities to improve scene perception in low-light and occluded environments. However, existing methods often struggle to balance perceptual quality and downstream task performance: approaches that produce visually pleasing fusion results may weaken semantic discriminability, while task-oriented methods can compromise visual fidelity. To address this issue, we propose SVSFusion, a visual-semantic-prior-guided self-prompting framework composed of three coordinated components. First, we introduce a geometry-aware 2D multi-directional scanning strategy to replace the conventional 1D token flattening adopted in existing state-space and Transformer-based models. By preserving spatial continuity during feature propagation, the proposed patch-wise recurrent architecture enables more effective long-range dependency modeling through gated bidirectional interactions. Second, a dual-branch prior extraction module is designed to generate complementary visual prompts, including semantic priors for infrared images derived from a pretrained segmentation network and structure-enhanced prompts for visible images obtained through a structure-tensor-guided enhancement branch. Third, a memory-gated spatial conditioning module constructs transient prompt-conditioned memory and selectively injects modality-specific hierarchical visual prompts into the fusion encoder, enabling controlled unidirectional prior transfer while reducing prior-induced feature degradation. Extensive experiments across multiple datasets demonstrate consistent improvements on perceptual metrics as well as in downstream tasks such as semantic segmentation and object detection. The proposed framework achieves a superior balance among visual quality, spatial structural integrity, and task adaptability. The code is available at: https://github.com/Gumpford/IVIF-SVSFusion .