AutoWeather4D: Autonomous Driving Video Weather Conversion via G-Buffer Dual-Pass Editing

Weather variation is critical for robust testing and validation of autonomous driving systems. Recent progress in generative editing has enabled weather manipulation in driving videos, but it often suffers from two key limitations. First, they exhibit physical inconsistency, such as unnatural light propagation and implausible dynamics of water or particles. Second, they offer limited control over fine-grained weather properties. To address these challenges, we propose AutoWeather4D, a novel framework that introduces a G-Buffer dual-pass editing mechanism: a Geometry Pass models interactions between weather elements and scene geometry, while a Light Pass adjusts illumination for day-night and weather-induced lighting changes. This dual-pass mechanism enables fine-grained control and generates physically plausible transitions. Experiments show that AutoWeather4D outperforms existing baselines in physical realism by avoiding ``impossible videos" and in controllability, generating diverse weather variations that closely align with real-world physics. This work advances autonomous driving simulation by providing a reliable tool for testing system robustness under complex weather conditions.