Factorized Learning for Temporally Grounded Video-Language Models

Wenzheng Zeng¹, Difei Gao¹, Mike Zheng Shou^1†, Hwee Tou Ng^1†,

¹National University of Singapore

^†Corresponding authors

ICCV 2025

📄 Paper with supp.   |   🤗 Daily paper   |   🤗 Model   |   🤗 Dataset   |     🖼️ Poster   |     ▶️ Video   |   X (Twitter)  |   知乎  

This repository contains the official implementation of the ICCV 2025 paper "Factorized Learning for Temporally Grounded Video-Language Models".

🔆 Highlights

• Model: We propose a new framework $D^2\mathrm{VLM}$, where we decompose the generation objective into a "grounding then answering with evidence referencing" paradigm and introduce evidence tokens to emphasize explicit event-level visual semantic capture.
• Training Algorithm: We introduce Factorized Preference Optimization (FPO) that explicitly addresses both temporal grounding and textual response. A factorized data synthesis approach is also designed to support FPO.
• Performance: Our method consistently outperforms SOTA methods across various tasks.
• Open Source: We release the source code and model weights to the community.

🔥News

• [2025-10] Code and model weights are released!
• [2025-06] Our work is accepted to ICCV 2025!

🛠️ Installation

Please refer to the following environmental settings that we use. You may install these packages by yourself if you meet any problem during automatic installation.

• CUDA 11.8
• Python 3.12.2
• PyTorch 2.4.0
• Transformers 4.44.2
• DeepSpeed 0.14.5
• NNCore 0.4.5

Install from source

1. Clone the repository from GitHub.

git clone https://github.com/nusnlp/d2vlm.git
cd d2vlm

2. Initialize conda environment.

conda create -n d2vlm python=3.12 -y
conda activate d2vlm

3. Install dependencies.

pip install -r requirements.txt

📦 Dataset

Please refer to the Dataset page.

🤖 Inference and Evaluation

1. You can download the pre-trained model at here.
2. Run the following commands. Remember to change the absolute path within each .sh file.

E.T. Bench

  bash scripts/inference.sh

  # or refer to the inference and Evaluation part of scripts/train_inference_eval.sh

Charades-STA

  bash all_benchmark_eval/charades/inference.sh

all_benchmark_eval/charades/inference.sh

Youcook2

  bash all_benchmark_eval/youcook2/inference.sh

💪 Training

1. Download the pretrained model from here (stage-2 model of E.T. Chat).
2. Check and run the following command (modify relevant path).

  bash scripts/train_inference_eval.sh

🎓 Citation

If you find our work useful in your research, please consider to cite our paper:

@inproceedings{d2vlm,
  title={Factorized Learning for Temporally Grounded Video-Language Models},
  author={Zeng, Wenzheng and Gao, Difei and Shou, Mike Zheng and Ng, Hwee Tou},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025},
  pages={20683-20693}
}

🙏 Acknowledgments

This project was built upon E.T. Bench, TimeChat, and AMP. We thank their solid contribution to the community!