Upsampling Digital Material Models Using Machine Learning for Radiance Accuracy

The system pre-trains a machine learning model on comprehensive training data that includes renderings at multiple light positions and resolutions. This preliminary training enables the model to learn the complex relationships between low-resolution inputs and high-resolution outputs under various lighting conditions, so that when actual upsampling is needed, the model can immediately apply this pre-learned knowledge to generate visually accurate results without performing computationally expensive real-time calculations.

The machine learning model acts as an intermediary between the low-resolution input digital material model and the desired high-resolution output. Instead of directly applying simple bilinear interpolation, the model processes the low-resolution input through learned transformations that account for lighting variations, effectively mediating the conversion to produce visually accurate high-resolution results that bilinear upsampling cannot achieve.

The system performs the computationally intensive model training phase in advance, during which the machine learning model learns to generate high-resolution digital material models that are visually realistic under varying lighting conditions. Once trained, the model can be deployed with relatively low computational overhead, as the heavy lifting of learning complex relationships has already been completed during the preliminary training phase.

The system creates a trained machine learning model that encapsulates the knowledge of how to generate visually realistic high-resolution digital material models. This model can then be copied and deployed across different systems or applications, allowing the computational complexity to be paid once during training rather than repeatedly during actual usage, thereby reducing the operational computational burden.