wyccccc/TimeChatOnline-7B
TimeChat-Online-7B by wyccccc is a 7 billion parameter online VideoLLM designed for efficient streaming video understanding, built upon the Qwen2.5-VL architecture. It features a Differential Token Drop (DTD) module that reduces visual redundancy by over 80% by preserving only meaningful temporal changes. This model excels at processing streaming video with high efficiency and proactive response capabilities, achieving new state-of-the-art results on benchmarks like StreamingBench and OVO-Bench.
Loading preview...
TimeChat-Online-7B: Efficient Streaming Video Understanding
TimeChat-Online-7B is an innovative 7 billion parameter online VideoLLM developed by wyccccc, specifically engineered for efficient processing of streaming video. Its core advancement is the Differential Token Drop (DTD) module, which intelligently identifies and discards redundant visual information in video streams, retaining only significant temporal changes. This mechanism allows the model to achieve over 80% token reduction without compromising performance, making it highly efficient for real-time applications.
Key Capabilities and Features
- Dynamic Video-aware Pruning: The DTD module adaptively reduces video tokens, suitable for various video speeds, from a holistic video perspective.
- Positional Reservation: It maintains precise spatial and temporal understanding by preserving fine-grained positions of retained tokens using M-ROPE.
- Streaming-friendly Design: DTD processes only newly arriving frames, avoiding re-processing historical content for efficient stream handling.
- Proactive Response: The model can autonomously detect critical scene transitions and moments that require responses, eliminating the need for explicit user queries.
- High Performance: TimeChat-Online-7B achieves state-of-the-art results on benchmarks such as StreamingBench (56.6 accuracy with 82.6% token reduction) and OVO-Bench (45.6 accuracy with 84.8% token reduction). It also significantly reduces video tokens (up to 85%) on long video benchmarks like MLVU, LongVideoBench, and VideoMME while maintaining or improving accuracy.
Ideal Use Cases
This model is particularly well-suited for applications requiring real-time analysis of streaming video, such as live monitoring, video surveillance, and interactive video assistants, where efficiency and the ability to focus on dynamic changes are critical.