Overview · Models and data · Codebase · Reproduction

Find the moments that matter.
TimeLens2 turns natural-language queries into precise, traceable evidence intervals on the video timeline.

_{One model family across seven temporal-grounding settings, from short indoor actions to long-form and egocentric video.}

---

Overview

TimeLens2 is a generalist video temporal-grounding MLLM. Given a video and a natural-language description or question, it finds when the supporting visual evidence occurs and returns one or more temporal intervals. A single model handles short and long videos, single and repeated events, descriptive and question-form queries, and both third-person and egocentric footage through a unified generative interface.

TimeLens2 treats temporal evidence as a set of intervals throughout training. Its supervised stage uses verified single- and multi-span annotations from TimeLens2-93K. Its GRPO stage combines temporal IoU with a matching-free temporal Wasserstein reward. The SFT, GRPO, and evaluation code are released in this repository.

Models and data

TimeLens2 uses two training stages. SFT learns evidence search and variable-cardinality interval generation; GRPO then calibrates the predicted intervals with temporal rewards.

The repository includes ready-to-use annotations and rollout data for the SFT and GRPO recipes. To reproduce the second stage, download the matching -SFT checkpoint and set it as MODEL_PATH for rollout generation or GRPO training. Video files are distributed separately through the linked Hugging Face dataset and must be downloaded before training. To use another framework for SFT or RL, download the public data and convert it to that framework's required format.

Codebase

The official training and evaluation code is organized around the three stages used in the project:

Reproduction

1. Prepare the videos. Download TimeLens2-93K, TimeLens-100K, and Ego4D-NLQ. The SFT and GRPO annotations are already included in this repository.
2. Run supervised fine-tuning. Launch train_sft_4b.sh or train_sft_8b.sh.
3. Prepare rollouts. Use the bundled results, or regenerate them with rollout_timelens2.sh. Its single configuration includes both timelens2-93k and timelens-100k.
4. Run GRPO. Launch train_grpo_4b.sh or train_grpo_8b.sh. Both default to the bundled official rollout files for both sources.
5. Evaluate. Run run_grounding.sh.

Each module has a focused README with installation, data layout, environment variables, and launch examples. Paths are configurable and may contain environment variables.

License and acknowledgements

The project is released under the Apache License 2.0. The SFT, evaluation, and GRPO modules contain code derived from InternLM/xtuner, open-compass/VLMEvalKit, and TencentARC/TimeLens, respectively. We thank their authors for open-sourcing these projects.

Citation

@misc{zhu2026timelens2,
      title={TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs},
      author={Yuhan Zhu and Changlian Ma and Xiangyu Zeng and Xinhao Li and Zhiqiu Zhang and Songze Li and Jun Zhang and Tianxiang Jiang and Yuandong Yang and Ziang Yan and Zikang Wang and Xinyu Chen and Haoran Chen and Shaowei Zhang and Limin Wang},
      year={2026},
      eprint={2607.17423},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.17423},
}