Brief

Researchers have developed OmniMem, a novel framework designed to improve the generation of long videos by efficiently retrieving relevant historical data. This method addresses the challenge of managing large key-value (KV) caches in autoregressive video models by employing sparse KV retrieval. OmniMem incorporates adaptive strategies to prevent local bias and manage memory access, allowing for more informative long-range retrieval without excessive memory overhead. Experiments demonstrate significant improvements in video consistency and detail preservation compared to existing techniques. AI