This paper presents real-time rendering using global illumination to provide realistic user experiences in virtual environments. We propose a learning-based estimator that predicts diffuse indirect illumination in screen space and combine it with direct illumination to synthesize a globally illuminated HDR result. It generalizes to handle complex lighting and scenarios and addresses the challenge of capturing long-range/long-distance indirect illumination when using neural networks. By applying a neural network approach to the rendering equation, we present a novel network architecture for predicting indirect illumination. It features a modified attention mechanism that aggregates global information guided by spatial geometric features and a monochromatic design that encodes each color channel individually. Experimental results demonstrate its superiority over existing learning-based techniques, demonstrating its ability to handle complex lighting conditions such as multi-colored lighting and environmental lighting. Furthermore, it successfully captures distant indirect illumination, simulates interreflections between textured surfaces (e.g., color bleeding), and effectively handles novel scenes not included in the training dataset.
