NeR-SC: Adapting Neural Video Representation to Screen Content

A new neural representation framework called NeR-SC has been proposed to improve the compression of screen content video, a category that includes remote desktop streams, online education recordings, and cloud gaming footage, according to a preprint published on arXiv [1]. The framework, detailed in a paper submitted on 26 May 2026, is built upon the SNeRV backbone and introduces three modules designed to exploit the distinct characteristics of screen content, which features sharp edges, limited color palettes, and strong temporal redundancy [1][2]. The first module is a learnable color palette that models the discrete color structure by restricting a low-frequency sub-band to a learned color set. The second is a multi-gate dense fusion module that replaces sequential feature fusion with dense, attention-gated cross-stage interaction. The third is an embedding-level frame skip strategy that bypasses redundant decoder invocations for static frames, adding zero training overhead [1][2]. In experiments on the DSCVC and VCD datasets, NeR-SC achieved average peak signal-to-noise ratio (PSNR) scores of 40.32 dB and 41.73 dB, respectively [1][2]. The authors report that the model outperforms representative neural video representation methods and, at low bitrates, surpasses the established H.264 and H.265 video compression standards [1][2]. The frame skip strategy also enables real-time decoding without any loss in quality, according to the paper [1][2]. Screen content video differs from natural video in several key ways. It often contains large uniform areas, repeated icons, and text, which existing neural representation methods, typically designed for natural scenes, do not efficiently exploit [2]. The development of NeR-SC addresses this gap by tailoring the representation to these specific statistics, potentially offering more efficient compression for applications where such content is dominant [1][2]. The paper appears on arXiv, a preprint server for scientific papers, and has not yet been peer-reviewed [1]. The research comes as the broader field of implicit neural representations continues to evolve as a paradigm for video compression, with methods increasingly competing with traditional codecs on specific types of content [1][2].