Scaling Zero-Shot Reference-to-Video Generation

TL;DR

Saber is a scalable zero-shot reference-to-video generation framework that learns from video-text pairs without explicit R2V data, using masked training and attention mechanisms to produce identity-consistent videos.

Contribution

Introduces Saber, a novel zero-shot R2V method that bypasses the need for explicit triplet data by leveraging video-text pairs and specialized training strategies.

Findings

Outperforms R2V-trained methods on OpenS2V-Eval benchmark.

Demonstrates strong generalization across multiple references.

Effectively reduces copy-paste artifacts with mask augmentation.

Abstract

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.