V-Bridge leverages the strong priors of generative models for image restoration tasks, achieving better performance while using only 0.1% of the training data required by conventional image restoration methods.
We use a single video generation model to support all image restoration tasks.
Each video shows three parts: Ours, FoundIR, and Ground Truth.
V-Bridge leverages video generative priors for versatile few-shot image restoration.
We pioneer the use of video generative models as universal priors, demonstrating that their inherent representations serve as a powerful, transferable foundation across diverse low-level tasks.
We propose V-Bridge, a data-efficient image restoration framework using progressive generative refinement and a coarse-to-fine training with lightweight drift correction.
V-Bridge
Experiments on multiple image restoration benchmarks show that V-Bridge matches or exceeds mainstream methods using only 0.1% to 5% of the training data.