StreamBridge: Turning Your Offline Video Large Language Model into a Proactive Streaming Assistant
This work solves the problem of enabling real-time, interactive video understanding for users of video AI systems, representing an incremental advancement by enhancing existing models with streaming features.
The paper tackles the problem of adapting offline Video-LLMs for streaming scenarios by addressing challenges in multi-turn real-time understanding and proactive responses, resulting in significant improvements in streaming capabilities that outperform proprietary models like GPT-4o and Gemini 1.5 Pro.
We present StreamBridge, a simple yet effective framework that seamlessly transforms offline Video-LLMs into streaming-capable models. It addresses two fundamental challenges in adapting existing models into online scenarios: (1) limited capability for multi-turn real-time understanding, and (2) lack of proactive response mechanisms. Specifically, StreamBridge incorporates (1) a memory buffer combined with a round-decayed compression strategy, supporting long-context multi-turn interactions, and (2) a decoupled, lightweight activation model that can be effortlessly integrated into existing Video-LLMs, enabling continuous proactive responses. To further support StreamBridge, we construct Stream-IT, a large-scale dataset tailored for streaming video understanding, featuring interleaved video-text sequences and diverse instruction formats. Extensive experiments show that StreamBridge significantly improves the streaming understanding capabilities of offline Video-LLMs across various tasks, outperforming even proprietary models such as GPT-4o and Gemini 1.5 Pro. Simultaneously, it achieves competitive or superior performance on standard video understanding benchmarks.