VMamba-ADD: A Dual-Branch Adversarial Diffusion Distillation Network for Efficient Blind Super-Resolution

Blind super-resolution aims to reconstruct high-resolution images from low-resolution inputs affected by complex and unknown degradations. Although adversarial diffusion distillation (ADD) has achieved significant improvements in inference speed, its aggressive distillation process inevitably weakens the modelling of long-range structures, leading to structural hallucinations and distorted details in complex scenes. To address these limitations, this study proposes VMamba-ADD. Specifically, a residual VMamba branch is incorporated into the U-Net decoder to efficiently capture global context. Meanwhile, an adaptive gating fusion module dynamically integrates global structural priors and local texture features. Furthermore, to suppress cross-step degradation propagation, we adopt the prediction-based self-refinement (PSR) strategy from the baseline ADD framework. Extensive experiments validate that the proposed method effectively suppresses generation artefacts and improves structural coherence. Quantitatively, VMamba-ADD improves the LPIPS metric from 0.3368 to 0.3366 under complex synthetic degradations and achieves a competitive MANIQA score of 0.6213 on the RealLR200 dataset, demonstrating robust perceptual fidelity. VMamba-ADD effectively mitigates the limitations of aggressive distillation. By generating structurally consistent high-resolution images while maintaining highly efficient inference (1–4 steps), it provides a reliable visual foundation for downstream engineering tasks, particularly automated infrastructure defect and crack segmentation.