In the field of Video Super-Resolution (VSR), accurately modeling inter-frame propagation and feature alignment are key to improving reconstruction quality. Although existing methods have achieved certain success, they still face challenges such as the accumulation of alignment errors, insufficient alignment for complex motions in dynamic scenes, and inadequate multi-frame feature fusion. This paper proposes a novel video super-resolution network called Selective Kernel and Offset Prediction Network (SK-OPNet). We designed two modules: the Multi-Input Selective Kernel Attention (M-SKA) module and the Multi-Scale Offset Prediction (MSOP) module. The M-SKA module dynamically fuses multi-frame features, enhancing information interaction across temporal scales and improving the stability and representational capacity of feature propagation. The MSOP module leverages multi-scale features for offset prediction, effectively increasing alignment accuracy in low-texture and large-motion regions. Experimental results show that SK-OPNet achieves excellent propagation stability and alignment performance, improving video super-resolution performance.