streaming video

VideoLLM runs live video QA at 2 FPS

A new VideoLLM system named AURA enables real-time live video question answering at 2 FPS, overcoming the limitations of prior models that processed only pre-recorded clips or struggled with continuous streaming. AURA achieves bounded latency by unifying a video encoder with an LLM and employing a sliding-window history with reusable prefix key-value caches.