Abstract: Existing convolutional neural network (CNN)-based loop filters tend to apply multiple networks for different quantization parameter (QP), consuming a significant amount of resources in training models and increasing memory burden. To address this issue, a CNN-based QP-adaptive loop filter is proposed. First, a lightweight classification network was designed to divide coding tree unit (CTU) into three categories—difficult, medium, and easy—based on the complexity of filtering. Furthermore, three CNN filtering networks, which incorporated feature information enhancement fusion modules, were constructed to meet the filtering needs of the three categories of CTU under different QP. The proposed loop filter was integrated into the versatile video coding (VVC) standard H.266/VVC test software VTM 6.0, replacing the original deblocking filter (DBF), sample adaptive offset (SAO), and adaptive loop filter. Experimental results show that the proposed method achieves an average reduction of 3.14% in Bjøntegaard delta bit rate (BD-BR), thereby significantly improving compression efficiency and reducing compression artifacts when compared to other CNN-based loop filters.