Adaptive Rendering via Linear Prediction Models

The patent segments the rendering process into two distinct phases: a tracing phase that uses linear prediction models for fast rendering of most pixels, and a refinement phase that applies Monte Carlo ray tracing only to selected positions where higher accuracy is needed. This segmentation allows the system to achieve high rendering quality for complex objects while dramatically reducing overall rendering time by avoiding exhaustive ray tracing across the entire image.

The patent implements local quality by adaptively determining which regions of the image require high-quality rendering versus fast rendering. Linear prediction models are used for regions where approximate rendering is sufficient, while Monte Carlo ray tracing is applied locally to specific positions that require higher accuracy. This local quality approach ensures that rendering resources are concentrated where they are most needed rather than uniformly applied across the entire image.

The patent applies partial action by using linear prediction models to render the majority of pixels in the image, accepting that these predictions are not perfectly accurate but are sufficiently good for most regions. This partial rendering approach covers most of the image with low computational overhead, while excessive action (full Monte Carlo ray tracing with many samples) is reserved only for specific positions where high accuracy is critically needed, thus balancing image quality with computational efficiency.

The patent introduces linear prediction models as an intermediary between the simple rendering approach and the computationally expensive Monte Carlo ray tracing. These prediction models serve as a mediator that provides reasonably accurate rendering for most pixels without requiring the full computational overhead of ray tracing, thereby reducing overall computational complexity while maintaining acceptable image quality.