♾️ Train Short, Inference Long: Training-free Horizon Extension for Autoregressive Video Generation

🌐 Project Page • 📄 arXiv

---

🎬 Long-Horizon Extension

FLEX enables stable long video generation beyond the predefined generation horizon — without retraining or finetuning.

Note: The GIF above is compressed for GitHub display and may suffer from reduced visual quality.
For higher-resolution videos and more demo, please visit Project Page: https://ga-lee.github.io/FLEX_demo/

---

✨ Overview

Autoregressive video diffusion models suffer from severe extrapolation failure, where rapid error accumulation leads to significant temporal degradation and motion collapse when extending beyond training horizons. we propose FLEX (Frequency-aware Length EXtension) , a training-free inference-time framework that bridges the gap between short-term training and long-term inference.

---

📊 Highlights

• 🔹 30s video generation (6× extrapolation for 5s models) outperforming SOTA opensource autoregressive methods
• 🔹 60s video generation (12× extrapolation for 5s models) competitive with long-video fine-tuned baselines
• 🔹 Training free, plug-and-play integration
• 🔹 Further support minute-level video generation

---

💻 Code

Code release is coming soon.
We are preparing a clean and well-documented implementation.

---

📄 Citation

If you find this work useful, please cite:

@misc{li2026trainshortinferencelong,
      title={Train Short, Inference Long: Training-free Horizon Extension for Autoregressive Video Generation}, 
      author={Jia Li and Xiaomeng Fu and Xurui Peng and Weifeng Chen and Youwei Zheng and Tianyu Zhao and Jiexi Wang and Fangmin Chen and Xing Wang and Hayden Kwok-Hay So},
      year={2026},
      eprint={2602.14027},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.14027}, 
}