FlexiVideo: Variation-Aware Temporal Dynamics Modeling for Efficient Video Understanding

Da Peng^1*, Xuesong Yang^2*, Zonghao Guo^3†, Yichen Zhang³, Chi Chen³,

Yidan Zhang², Yuan Yao³, Fang Wan^2†, Wei Ke^1†, Maosong Sun³,

¹Xi'an Jiaotong University, ²University of Chinese Academy of Sciences, ³Tsinghua University

* Equal contribution † Corresponding author

Accepted to CVPR 2026!

🌟 What is FlexiVideo?

Natural videos exhibit heterogeneous temporal dynamics, with certain segments undergoing high-dynamic scene transitions and others dominated by low-dynamic visual changes. However, treating all frames identically, a common practice in most MLLMs, leads to redundant visual encoding, which results in significant computational overhead. The recent SoTA model, i.e., Qwen2.5-VL, adopts a fixed two-frame encoding scheme, but our pilot experiments indicate that it encounters a visual confusion problem under high-dynamic frame pairs. To address this issue, we propose FlexiVideo, an efficient MLLM that models temporal dynamics leveraging visual variation. FlexiVideo first employs an adaptive temporal segmentation module to estimate inter-frame differences, grouping consecutive frames into scene segments with subtle visual changes. Subsequently, a dynamical spatio-temporal embedding module adjusts the temporal window for scene-level encoding. By restructuring scene-level visual representations within a structured temporal organization, our approach models dynamics more effectively and reduces the encoding burden while preserving fine-grained visual variations. Extensive experiments show that FlexiVideo consistently outperforms Qwen2.5-VL-3B across 6 general video benchmarks. Notably, when evaluated on MotionBench at 10 FPS, FlexiVideo reduces visual tokens by 43.5% compared with Qwen2.5-VL-3B while achieving a 1.3% performance gain, striking a significantly better balance between efficiency and effectiveness.

🔥 News

• [2026/02/21] 📢 FlexiVideo is accepted by CVPR2026.

🚀 Quick Start

Set up a new virtual environment

conda create -n flexivideo python=3.10 -y
conda activate flexivideo
bash build_env.sh

Inference

python run.py # training-free on Qwen2.5-VL 

Evaluation

cd VLMEvalKit 
bash eval.sh # training-free on Qwen2.5-VL 

---

🧩 To-Do List

• Release the Inference Scripts
• Release the Evaluation Scripts
• Release the Training Scripts
• Release the Training Data

---

This repo benefits from LlamaFactory and VLMEvalKit. Thanks for their wonderful works.

📬 Contact

For any questions or collaborations, feel free to contact us : )

📧 MetaPDa@gmail.com

📖 Citation

If you find FlexiVideo useful in your research, please consider citing:

@inproceedings{peng2026flexivideo,
  title={FlexiVideo: Variation-Aware Temporal Dynamics Modeling for Efficient Video Understanding},
  author={Peng, Da and Yang, Xuesong and Guo, Zonghao and Zhang, Yichen and Chen, Chi and Zhang, Yidan and Yao, Yuan and Wan, Fang and Ke, Wei and Sun, Maosong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9804--9814},
  year={2026}
}