PSL-Flow: Physics-Structured Latent Flow Matching for Aerial Visible-to-Infrared Image Translation

Abstract

Aerial visible-to-infrared (V2IR) image translation provides scalable data support for nighttime remote sensing. However, existing methods predominantly learn cross-modal mappings within pixel or generic latent spaces, where physical information acts merely as an auxiliary training constraint. This limitation hinders physical cues from continuously participating in thermal structure generation during inference. In this letter, we propose PSL-Flow, a novel framework for aerial V2IR image generation via conditional flow matching within a physics-structured latent space (PSL-VAE). PSL-Flow shifts the generative objective from direct pixel mapping to physical quantity distribution modeling, enabling physical information to serve as intrinsic generative variables. Specifically, physical quantities including temperature are estimated from real infrared (IR) images to construct an intermediate representation endowed with thermal imaging semantics. Subsequently, the physical backbone and local residuals are jointly encoded into a PSL-VAE, transforming the objective from pixel mapping to thermal state modeling. Ultimately, a continuous generation process driven by visible conditions is learned within this latent space, and IR images are reconstructed through the synergistic decoding of the physical backbone and boundary residuals. Experimental results demonstrate that the proposed method achieves stable advantages across multiple aerial datasets. Our code is available at https://github.com/zbZ113/PSL-Flow