Haoran Bai

SATB-VR: Training Few-Step Video Restoration Diffusion Model using SNR-Aware Trajectory Blending

Haoran Bai, Xiaoxu Chen, Xiaoyu Liu, Zongsheng Yue, Sibin Deng, Wangmeng Zuo, et al.

While diffusion models excel in video restoration, their reliance on extensive iterative steps limits efficiency. Conversely, aggressive single-step distillation often compromises fine texture recovery. To achieve an optimal balance, we present SATB-VR, a few-step paradigm that j…Read more