Realistic infrared image generation based on physics-guided latent diffusion
Huanyu Yang, Mengchu Tian, Jun Wang, Jun Wang, Yuming Bo, Jiacun Wang, Jiacun Wang, Henry Han, Peng Zhu, Giancarlo Fortino · Engineering Applications of Artificial Intelligence · 2026
Infrared image generation is essential in scenarios with low illumination or complex environments, but the scarcity of aligned visible–infrared data and the lack of physical realism in generated results remain key challenges. However, existing generative models often overlook the thermodynamic principles underlying infrared imaging, resulting in synthetic images that are visually plausible yet physically inaccurate. In this paper, we propose Infrared Physics-guided Latent Diffusion (IPLD), a novel framework that integrates physics-guided modeling into a latent diffusion process for high-fidelity synthesis of infrared images. Central to IPLD is the Temperature–Emissivity–Environmental Radiance (TeR) decomposition, which decomposes thermal signals into temperature, emissivity, and environmental radiance components, governed by the laws of blackbody radiation. To enhance the environmental radiance modeling, we introduce Environmental Radiance Map Estimation (ERME), a hybrid local–global estimation mechanism that preserves both spatial detail and thermal consistency. Furthermore, a novel Skip Connection Diffusion Transformer (SCDT) is proposed to strengthen and balance semantic structure and fine-grained details during image reconstruction. Extensive experiments on public datasets demonstrate that IPLD outperforms state-of-the-art Generative Adversarial Network (GAN)-based and diffusion-based methods, achieving superior results in Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS), and Fréchet Inception Distance (FID) metrics. Ablation studies validate the complementary value of TeR, ERME, and SCDT in improving radiative realism. Our approach establishes a new paradigm for physically grounded image translation, offering enhanced generalization and reliability for downstream perception tasks such as target detection.