Froxel Grid Upscaling for Temporally Coherent Volumetric Rendering

The patent creates a training set where high-resolution volumetric renderings are generated by repeatedly rendering the same frozen scene state at high resolution, then uses these pre-computed high-res images to train a machine learning model. The model learns to generate high-resolution volumetric effects from low-resolution inputs, effectively copying the complex rendering results without performing the expensive computations in real-time during gameplay.

The patent performs the computationally intensive high-resolution rendering in advance during a training phase, before the actual gameplay occurs. By pre-computing and storing high-resolution volumetric data for various scene states, the system eliminates the need to perform these expensive calculations during real-time gameplay, thus reducing computational burden while maintaining high image quality.

The patent uses machine learning to copy and upscale low-resolution volumetric data into high-resolution output. The trained model takes low-resolution frame data as input and generates high-resolution output that maintains temporal coherence, effectively copying the quality of high-resolution rendering while operating at low-resolution computational cost during real-time gameplay.

The patent changes the resolution parameter dynamically - using low resolution for computational efficiency during gameplay while the machine learning model compensates by generating high-resolution output through upscaling. This parameter transformation allows the system to maintain both high frame rates and high temporal resolution without the traditional trade-off.

The patent trains the machine learning model to copy high-resolution volumetric characteristics directly from pre-computed training data. Instead of using temporal blending to smooth results, the model learns to preserve temporal coherence by copying the precise high-resolution details that were captured during the training phase, eliminating the need for post-processing blending operations.

The patent replaces the mechanical approach of temporal blending (which smooths but blurs temporal details) with a machine learning-based upscaling approach. The neural network model substitutes for the traditional image processing pipeline, generating sharp high-resolution output that maintains temporal coherence without relying on blending operations that would compromise temporal stability.