Generative Video Compression

Generative video compression studies how to encode a small amount of information and use learned models to reconstruct video, reducing bitrate by synthesizing details that conventional codecs would transmit. Methods include compact latent and token codecs, sparse keyframes or motion cues, autoregressive context models, and diffusion or flow-based decoders; some use pretrained models without task-specific training, while others train or distill models for compression. On benchmarks such as UVG, HEVC, and MCL-JCV, work often compares perceptual metrics such as LPIPS and DISTS alongside bitrate, temporal consistency, and decoding speed, especially at ultra-low bitrates. The central tension is how to balance perceptual realism and compactness against faithful content, coherent motion, and the compute and latency required for practical streaming.