Neural representation for videos (NeRV) has emerged as a promising method for video representation and compression. However, existing NeRV methods primarily focus on objective quality and overlook subjective quality. Considering the varying sensitivity of human eyes to different regions, we propose a saliency-based neural representation for videos (SNeRV). By introducing a multi-scale temporal-spatial feature grid and SNeRV blocks, we enhance the model’s representation capability, improving both objective and subjective quality. Additionally, our saliency-guided training strategy enables more efficient parameter allocation, prioritizing the representation of regions of interest (ROI) for superior visual quality. On the UVG dataset, our proposed method improves objective quality by 0.3 dB to 0.5 dB PSNR compared to the state-of-the-art method and significantly enhances subjective quality, particularly in ROI areas.