Neural Video Compression
Overview
Neural video compression studies how to encode video with fewer bits while keeping reconstructed frames useful and consistent over time. Common approaches include motion-compensated codecs that predict frames from references and entropy-code the remaining information, implicit neural representations that encode a video as a compact model, and generative codecs that use learned priors to synthesize detail at very low bitrates. Methods are compared on benchmarks such as UVG and HEVC using measures of pixel fidelity and perceptual quality, alongside practical concerns such as decoding speed, memory, and robustness across devices. A central tension is how to gain bitrate savings without sacrificing exact detail or temporal stability, especially when generative reconstruction can invent content and may require substantial computation.
Latest papers
53 papers4 papers in the last four weeks, down 60% on the four weeks before. 0.0% of all new papers.