Researchers have developed SGVC, a novel streaming generative video compression framework designed to improve efficiency and quality at ultra-low bitrates. SGVC adapts a pretrained latent video diffusion transformer (DiT) into a causal pixel-space refiner, projecting decoded features into the DiT's embedding space. To bridge representation gaps, it employs Embedding Alignment during training and In-Context Propagation to reuse layer-wise states for temporal consistency without buffering. Experiments show SGVC achieves significant rate-perception performance, outperforming S$^2$VC by an average of 71.97% in DISTS. AI
IMPACT This research could lead to more efficient video streaming and storage solutions by improving generative video compression techniques.
RANK_REASON This is a research paper detailing a new method for video compression. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →