Reference-to-video (R2V) generation aims to synthesize videos that align with a text prompt while preserving the subject identity from reference images. However, current R2V methods are hindered by the reliance on explicit reference image-video-text triplets, whose construction is highly expensive and difficult to scale. We bypass this bottleneck by introducing Saber, a scalable zero-shot framework that requires no explicit R2V data. Trained exclusively on video-text pairs, Saber employs a masked training strategy and a tailored attention-based model design to learn identity-consistent and reference-aware representations. Mask augmentation techniques are further integrated to mitigate copy-paste artifacts common in reference-to-video generation. Moreover, Saber demonstrates remarkable generalization capabilities across a varying number of references and achieves superior performance on the OpenS2V-Eval benchmark compared to methods trained with R2V data.
Cite
@article{arxiv.2512.06905,
title = {Scaling Zero-Shot Reference-to-Video Generation},
author = {Zijian Zhou and Shikun Liu and Haozhe Liu and Haonan Qiu and Zhaochong An and Weiming Ren and Zhiheng Liu and Xiaoke Huang and Kam Woh Ng and Tian Xie and Xiao Han and Yuren Cong and Hang Li and Chuyan Zhu and Aditya Patel and Tao Xiang and Sen He},
journal= {arXiv preprint arXiv:2512.06905},
year = {2025}
}