Induces pretrained video models to perform SOTA image restoration using less than 2% of the training data required by specialized architectures.
arXiv · March 16, 2026 · 2603.13089
Why it matters
This proves that video generative models implicitly learn powerful, transferable restoration priors that can be unlocked with just 1,000 multi-task samples, suggesting a new 'foundation model' approach for low-level vision.
From the abstract
Large-scale video generative models are trained on vast and diverse visual data, enabling them to internalize rich structural, semantic, and dynamic priors of the visual world. While these models have demonstrated impressive generative capability, their potential as general-purpose visual learners remains largely untapped. In this work, we introduce V-Bridge, a framework that bridges this latent capacity to versatile few-shot image restoration tasks. We reinterpret image restoration not as a sta