Abstract: Although deep learning-based models can effectively handle multi-modal feature fusion, they often suffer from detail loss when compensating visible light with infrared in low-light environments, which limits further improvement of image quality. To address this issue, this paper proposes a CNN-Transformer dualbranch feature-driven fusion network (DFDFuse). The encoding stage employs a dual-branch progressive feature processing mechanism: specifically, a combination of reversible neural networks and convolution operations handles local features, while the Avgformer (with attention computation replaced by average pooling) processes global low-frequency information. This design enhances the model's focus on target details and improves cross-modal detail extraction. To strengthen the expression of salient features and suppress noise interference, we introduce a Differential Feature Extraction (DFE) module and a Fusion Features Enhancement (FFE) in the fusion layer to amplify differentiated detail information. Extensive experiments on multiple datasets demonstrate that our method effectively preserves visible-light texture details while maintaining thermal radiation target information. The proposed approach comprehensively outperforms CDDFuse, achieving approximately 10% improvement across seven qualitative evaluation metrics.