Adverse Weather Removal via Dynamic Enhancement Diffusion With Weather-Adaptive Prompting

Adverse weather conditions, such as rain, fog, and snow, degrade visual information, posing significant challenges for image restoration frameworks that adapt across diverse scenarios. Unified models for weather removal often struggle to capture weather-specific details, while two-stage methods require additional training and prior weather knowledge, limiting integration efficiency. In this paper, we propose a novel plug-and-play framework that enables unified models to achieve weather-aware enhancement without relying on weather priors. Our approach consists of two key components: 1) a weather-adaptive prompt that establishes weather awareness by generating transition features to quantify the enhancement needed for transitioning from adverse to clear conditions, guided by a transition knowledge generation (TKG) loss; and 2) a dynamic enhancement diffusion model that leverages these weather-aware transition features to iteratively restore images, adapting enhancement to each specific condition through a learned weather-variant distribution. Comprehensive experimental results across varied weather scenarios validate the effectiveness of our method.