PhysRain: Controllable Rainy Driving Scene Synthesis with Physics-Based Rain Rendering

Abstract

Text-to-image models can synthesize visually plausible rainy traffic scenes, but text prompts provide only coarse appearance-level control. They cannot explicitly specify or monotonically regulate measurable rain structures such as streak density, wind-driven orientation, exposure-dependent length, and foreground occlusion. To make rainy-scene synthesis physically controllable, we present PhysRain, a physics-guided framework that uses a physics-based renderer to provide structural rain priors for generative image editing. The renderer consists of two main components: a raindrop distribution, motion, and projection module that models the 3D rain field from interpretable weather, depth, exposure, and camera parameters; and an appearance modeling module that converts projected rain trajectories into structural masks with attenuation, defocus blur, and streak kernels. Experiments on Cityscapes show highly consistent physical control: for rain-rate control, mask coverage and mass both reach Spearman correlation 0.9989; for wind-speed control, mean absolute streak angle reaches Spearman correlation 1.0000; and for wind-direction control, east-west polarity accuracy is $\mathbf{1. 0 0 0 0}$ with $\mathbf{0. 9 8 2 0}$ signed-angle Spearman correlation.