This contribution presents a lightweight super-resolution filter that combines convolutional neural networks (CNNs) with existing RPR functionality in VVC. In this contribution, we modify the basic unit of MMSDAB proposed by JVET-AA0065 and add spatial attention into MMSDAB to make MMSDANet lightweight, called LMSDANet. Compared with VTM-11.0_NNVC-2.0, the proposed CNN filter achieves average {-9.16%, 17.03%, -7.61%} and {-4.14%, 6.34%, -2.25%} BD-rate gains (average on A1 and A2) in AI and RA configurations, respectively.