Search papers, labs, and topics across Lattice.
IR-Flow leverages Rectified Flow to create a unified framework for image restoration, bridging the gap between single-step discriminative methods and multi-step generative approaches. It constructs multi-level data distribution flows and learns transport trajectories using cumulative velocity fields, guided by a multi-step consistency constraint. Experiments on deraining, denoising, and raindrop removal show that IR-Flow achieves competitive results with few sampling steps while maintaining a good distortion-perception balance.
Linear transport flows between degraded and clean image domains enable fast, adaptable image restoration that outperforms existing methods in distortion-perception balance.
In image restoration, single-step discriminative mappings often lack fine details via expectation learning, whereas generative paradigms suffer from inefficient multi-step sampling and noise-residual coupling. To address this dilemma, we propose IR-Flow, a novel image restoration method based on Rectified Flow that serves as a unified framework bridging the gap between discriminative and generative paradigms. Specifically, we first construct multilevel data distribution flows, which expand the ability of models to learn from and adapt to various levels of degradation. Subsequently, cumulative velocity fields are proposed to learn transport trajectories across varying degradation levels, guiding intermediate states toward the clean target, while a multi-step consistency constraint is presented to enforce trajectory coherence and boost few-step restoration performance. We show that directly establishing a linear transport flow between degraded and clean image domains not only enables fast inference but also improves adaptability to out-of-distribution degradations. Extensive evaluations on deraining, denoising and raindrop removal tasks demonstrate that IR-Flow achieves competitive quantitative results with only a few sampling steps, offering an efficient and flexible framework that maintains an excellent distortion-perception balance. Our code is available at https://github.com/fanzh03/IR-Flow.
