TGFusion: Text-Guided Flow Matching for Image Fusion Under Complex Degradations

Researchers have introduced TGFusion, a text-guided latent-space flow matching framework designed to address the challenges of infrared-visible image fusion under realistic degradation scenarios. The framework unifies degradation suppression and cross-modal fusion by encoding task, degradation, and generation cues into structured prompts. Unlike existing methods that inject fixed global text representations, TGFusion adapts textual guidance to spatially varying degradations, local structures, and thermal saliency. The approach leverages text to provide prior information about degradation characteristics, complementing limited evidence from corrupted input images. The paper, available on arXiv (2608.00530), proposes a Prompt-conditioned Multi-stream J (likely a typo for 'Joint') architecture to fully exploit these priors. This work is significant for computer vision applications where image fusion must operate under complex, real-world conditions.