TeleBoost

A systematic alignment framework for high-fidelity, controllable, and robust video generation

Yuanzhi Liang, Xuan'er Wu, Yirui Liu,
Yijie Fang, Yizhen Fan, Ke Hao, Rui Li, Ruiying Liu, Ziqi Ni, Peng Yu, Yanbo Wang,
Haibin Huang, Qizhen Weng, Chi Zhang, Xuelong Li
Institute of Artificial Intelligence, China Telecom (TeleAI)
Paper Code (Coming soon) Contact

Abstract

Post-training is the critical step that transforms a pretrained video generator into a production-ready model that follows instructions, remains controllable, and stays stable over long horizons. TeleBoost organizes supervised policy shaping, reward-driven reinforcement learning, and preference-based refinement into a stability-constrained optimization stack. Designed for real-world video-generation constraints—high rollout cost, temporally compounding failure modes, and heterogeneous, uncertain, often weakly discriminative feedback—the stack treats optimization as a staged, diagnosis-driven process to improve perceptual fidelity, temporal consistency, and prompt adherence while preserving controllability established at initialization.

Key Contributions

  • Systematic post-training stack combining supervised shaping, reward RL, and preference refinement under stability constraints.
  • Design tailored to high-cost rollouts, temporal compounding errors, and heterogeneous/uncertain feedback signals.
  • Staged, diagnosis-driven optimization that lifts fidelity, temporal consistency, and prompt following while retaining controllability.
  • Unified recipe for robust, high-fidelity video generation ready for production deployment.

Comparisons

Side-by-side comparisons across methods. Each case shows Baseline, DanceGRPO, and TeleBoost.

Image-to-Video (I2V)

Jump to T2V →

Text-to-Video (T2V)

Back to I2V ↑

Citation


            @misc{liang2026teleboostsystematicalignmentframework,
                    title={TeleBoost: A Systematic Alignment Framework for High-Fidelity, Controllable, and Robust Video Generation}, 
                    author={Yuanzhi Liang and Xuan'er Wu and Yirui Liu and Yijie Fang and Yizhen Fan and Ke Hao and Rui Li and Ruiying Liu and Ziqi Ni and Peng Yu and Yanbo Wang and Haibin Huang and Qizhen Weng and Chi Zhang and Xuelong Li},
                    year={2026},
                    eprint={2602.07595},
                    archivePrefix={arXiv},
                    primaryClass={cs.CV},
                    url={https://arxiv.org/abs/2602.07595}, }