PVRF uses VLMs and rectified flow to remove rain, snow, and fog from images

A team of researchers (Wei Dong, Han Zhou, Terry Ji, et al.) has released PVRF, a new framework that tackles the long-standing challenge of removing heterogeneous adverse weather conditions (rain, snow, fog) from images without needing labeled training data for each condition. Traditional methods either rely on distortion-driven training that yields overly smooth results or fail to generalize to unseen combinations of degradations. PVRF solves this by integrating two novel components: a zero-shot weather perception module and a velocity-constrained rectified flow refinement.

The perception module uses frozen vision-language models (VLMs) in an AWR-specific question-answering format (AWR-QA) to estimate soft probabilities of weather types and low-level attribute scores. These estimates then condition a restoration network via attribute-modulated normalization (AMN) and weather-weighted adapters (WWA), producing an initial anchor estimate. To refine this estimate, the model learns a terminal-consistent residual rectified flow with perception-adaptive source perturbation and a terminal-consistent velocity parameterization that stabilizes learning near the terminal regime. In extensive experiments on single and combined degradations, PVRF shows consistent improvements in both fidelity (e.g., PSNR) and perceptual quality (e.g., LPIPS) over state-of-the-art baselines, with strong generalization across datasets. Code is promised for release.