RectiWeather: Photo-Realistic Adverse Weather Removal via Zero-shot Soft Weather Perception and Rectified Flow

Wei Dong, Han Zhou, Terry Ji, Guanhua Zhao, Yulun Zhang, Guangtao Zhai, Xiaohong Liu, Jun Chen

Keywords: zero-shot, soft perception, rectified flow

Abstract: Despite significant progress in Adverse Weather Removal (AWR), challenges remain in applying existing methods to real-world scenarios and in generating photo-realistic and visually compelling outcomes. The limited generalization of current approaches can be attributed to their inability to accurately perceive complex degradations in weather-affected images. Moreover, owing to optimization objectives that prioritize distortion losses, discriminative methods often produce overly smooth reconstructions. To address these challenges, we propose \textbf{RectiWeather}, a novel AWR framework guided by zero-shot soft perceptions extracted from pre-trained vision–language models (VLMs). Specifically, we design an AWR-specific Question Answering (AWR-QA) module that guides VLMs to produce soft perceptions of weather conditions and low-level attributes. These soft perceptions are then integrated into baseline AWR models through attribute-modulated normalization (AMN) and weather-weighted adapters (WWA), enabling posterior mean estimation while minimizing distortion loss. Furthermore, we map the posterior output to the clean image distribution using a perception-aware rectified flow model, where soft perceptions define the source distribution and guide the velocity field. Extensive experiments show that RectiWeather consistently surpasses state-of-the-art baselines in fidelity and perceptual metrics across both all-in-one and out-of-distribution scenarios. Our code will be released upon publication.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 60