TSDeblur: A Long-Range Temporal-Spatial Framework for Video Deblurring

Video deblurring remains challenging due to the difficulty of modeling long-range temporal dependencies and global spatial structures in a stable and efficient manner. Existing temporal approaches typically rely on short-frame windows or implicit recurrent propagation, both of which struggle to maintain controllable information flow over extended sequences and often accumulate errors. Spatially, global modeling with convolutions or attention incurs high computational costs, limiting efficiency. We address these limitations with TSDeblur, a temporal–spatial video deblurring framework derived from an optimization-inspired perspective. First, we introduce an explicit video degradation formulation that incorporates adjacent-frame information, enabling more transparent and controllable temporal modeling. Based on this formulation, we construct a proximal gradient descent (PGD)-inspired unfolding architecture. The gradient-descent step is implemented with a lightweight inter-frame fidelity module that enforces data consistency using nearby frames, while the proximal mapping step adopts a recurrent structure to establish long-range temporal dependencies. For spatial modeling, we integrate frequency-domain convolution into the proximal mapping stage, leveraging the global receptive field of Fourier coefficients to capture long-range spatial structures in a lightweight and computationally efficient manner. Extensive experiments on synthetic and real-world datasets demonstrate that TSDeblur achieves competitive performance among state-of-the-art methods while maintaining a compact architecture and low inference cost.