3D Rendering Neural Network for Sparse Sampling

The patent segments the rendering process into two distinct stages: (1) a training phase where a neural network is trained using rendering results from a dense set of sample points, and (2) a rendering phase where the trained neural network generates shading effects for a sparse set of sample points. This segmentation allows the computationally intensive dense sampling to be performed only once during training, while subsequent rendering operations use the trained model to efficiently process sparse samples, thereby resolving the contradiction between rendering accuracy and rendering speed.

The patent replaces the traditional mechanical/computational approach of performing complex shading calculations (such as ray tracing or radiosity) for all points with a neural network-based system. The neural network is trained to learn the complex shading relationships from a training set, and then during rendering, it directly predicts shading effects for sparse sample points without performing the original complex calculations. This substitution dramatically reduces computational complexity while preserving shading accuracy.

The patent performs preliminary action by training a neural network model in advance using rendering data from a comprehensive set of sample points. This training process captures the complex relationships between scene geometry, lighting, and shading effects. Once trained, the model can accurately predict shading effects for sparse sample points during actual rendering, thereby maintaining high image quality while achieving rendering efficiency through the use of minimal sample points.