DPAFuse: A dual-space probabilistic adversarial image fusion framework with robust coding embedding.

Multimodal image fusion aims to integrate complementary information from heterogeneous sensors, yet existing methods remain vulnerable to composite degradations and inherent modality bias. On the one hand, representations become unreliable under composite degradations; on the other hand, fusion strategies often directly enforce cross-modal consistency, inevitably forcing the fused results into inter-modal compromises and leading to unstable representations. To address these issues, we propose DPAFuse, a dual-space probabilistic adversarial image fusion framework with robust coding embedding. Firstly, a robust coding embedding mechanism projects heterogeneous inputs into a unified latent space, contracting the distributional discrepancies among diverse composite degradations and aligning them closer to the underlying clean representation, while effectively suppressing degradation-related components and preserving modality-invariant structural and semantic features. Moreover, to alleviate modality bias, we design a dual-space probabilistic adversarial fusion mechanism that enforces distribution-level consistency across latent and image domains, enabling adaptive global fusion without local compromises. Extensive experiments on public datasets demonstrate that DPAFuse achieves superior robustness and fusion fidelity compared with state-of-the-art methods. The code is publicly available at https://github.com/HaoZhang1018/DPAFuse.