MCG-NJU/TimeLens2-8B
MCG-NJU/TimeLens2-8B is an 8 billion parameter video multimodal large language model developed by MCG-NJU, designed for temporal grounding tasks. Built upon Qwen3-VL-8B-Instruct, it identifies precise time intervals within a video corresponding to a given text query. This model achieves a 48.0 average mIoU across seven temporal grounding benchmarks, setting a new state-of-the-art for accurately localizing events in short, long, and egocentric videos.
Loading preview...
TimeLens2-8B: Video Temporal Grounding Model
TimeLens2-8B is a specialized video multimodal large language model (VM-LLM) developed by MCG-NJU, focusing on temporal grounding. This capability allows the model to pinpoint specific time segments within a video that are relevant to a user's text query.
Key Capabilities & Features
- Precise Temporal Grounding: Given a video and a natural language query (e.g., "A man opens the refrigerator."), TimeLens2-8B returns a JSON array of
[start, end]time intervals in seconds where the described event occurs. - Strong Performance: The model achieves an impressive 48.0 average mIoU (mean Intersection over Union) across a suite of seven diverse temporal grounding benchmarks.
- State-of-the-Art: TimeLens2-8B has established a new state-of-the-art performance record on this seven-benchmark suite, demonstrating robust capabilities across various video types, including short, long, and egocentric footage.
- Foundation Model: It is built upon the powerful Qwen3-VL-8B-Instruct architecture, leveraging its multimodal understanding.
Ideal Use Cases
TimeLens2-8B is particularly well-suited for applications requiring accurate event localization within video content, such as:
- Video Search and Retrieval: Quickly finding specific moments in long videos based on textual descriptions.
- Content Moderation: Identifying and timestamping objectionable content.
- Video Summarization: Automatically highlighting key events.
- Assisted Video Editing: Helping editors locate relevant clips efficiently.