Using GLO-GAN for Realistic Lighting and Shadow Generation in Urban Art Visualization

To address the difficulty in balancing the physical realism of light and shadow with the representation of region-specific chromatic distribution, surface ornament regularity, façade composition patterns, and material reflectance characteristics in urban art visualization, this paper proposes a Generative Latent Optimization–Generative Adversarial Network (GLO-GAN) architecture driven by physical constraints and cultural perception. The framework refines the GLO-GAN architecture by incorporating material-aware attributes to enhance the consistency between synthesized urban scenes and their cultural context. A dual-branch generator with a shared GLO-optimized latent space is employed to jointly model physical illumination and semantic characteristics. Differentiable physical lighting constraints, including sun trajectories, ambient occlusion, and material reflectance, are integrated to ensure physically plausible rendering, while urban cultural labels provide semantic guidance for region-aware feature generation. Joint optimization combining L2 reconstruction, VGG perceptual, adversarial, and lighting consistency losses enables coordinated feature learning. A coarse-to-fine progressive upsampling strategy together with a multi-scale PatchGAN discriminator suppresses artifacts and enhances local details, whereas the StyleGAN3 backbone guarantees alias-free feature propagation and stable spatial phase alignment during illumination modulation. Considering that physically consistent light–material interaction is also fundamental to electromagnetic wave propagation analysis and computational imaging, the proposed framework provides methodological insights for intelligent electromagnetic sensing and multimodal scene perception. Experimental results demonstrate high performance, achieving a solar azimuth RMSE of 4.2◦, cultural style classification accuracy of 92.7%, and an average SSIM of 0.863. The proposed method enables realistic urban art visualization while preserving stable chromatic tendencies, structural motifs, and material-associated tonal distributions, and further offers a practical computational paradigm for integrating heterogeneous material information into intelligent electromagnetic-aware visualization systems.