Machine Learning Image Illumination Rendering for Consistent Frame Rates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Ray tracing for image rendering is computationally expensive and varies unpredictably, making it difficult to maintain a consistent frame rate and quality in image rendering processes.

Innovation Solution

Employ a machine learning system, such as a neural network, to learn the relationship between pixel surface properties and rendered pixels, approximating ray-traced images with a consistent computational budget by training on high-quality ray-traced images to predict pixel values based on material properties and viewing angles.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If ray tracing is used for image rendering, then image quality is improved, but computational cost increases and frame rate consistency deteriorates

Engineering Contradiction:
Improveimage qualityVSAvoidframe rate consistency
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent pre-calculates and stores illumination data for different lighting conditions and material properties before runtime. During actual rendering, the system queries this pre-computed data instead of performing full ray tracing, thereby maintaining high image quality while achieving consistent frame rates.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates simplified copies of complex illumination data in the form of lookup tables and pre-computed radiance values. These copies approximate the results of full ray tracing but can be retrieved and applied much faster, resolving the contradiction between quality and speed.

Inventive Principle:
Principle #26Copying

2Productivity

If average image quality is set low, then frame rate consistency is improved, but image quality deteriorates

Engineering Contradiction:
Improveframe rate consistencyVSAvoidimage quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent changes the parameters used for illumination calculation by switching between different levels of pre-computed data depending on the specific rendering needs. This allows the system to maintain high quality where needed while achieving consistent performance across varying scenes.

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If average image quality is set close to maximum, then image quality is improved, but frame rate consistency deteriorates

Engineering Contradiction:
Improveimage qualityVSAvoidframe rate consistency
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent implements dynamic selection of rendering techniques based on scene characteristics. The system automatically chooses between pre-computed illumination data and more detailed calculations based on the specific lighting conditions, material properties, and viewport requirements, thereby maintaining maximum quality while ensuring consistent frame rates.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4095807B1Image illumination rendering using a plurality of machine learning models
Publication Date: 2025.07.02 SONY INTERACTIVE ENTERTAINMENT LLC
  • EP4095807B1 patent drawingFigure 1
  • EP4095807B1 patent drawingFigure 2
  • EP4095807B1 patent drawingFigure 3

AI summary

An image rendering method for a virtual scene comprises, for a plurality of IDs, generating a respective mask identifying elements of the scene that are associated with a respective ID; for the resulting plurality of masks, dividing a respective mask into a plurality of tiles; and discarding tiles that do not identify any image elements; for the resulting plurality of remaining tiles, selecting a respective trained machine learning model from among a plurality of machine learning models, the respective machine learning model having been trained to generate data contributing to a render of at least a part of an image, based upon elements of the scene associated with the same respective ID as the elements identified in the mask from which the respective tile was divided; and using the respective trained machine learning model to generate data contributing to a render of at least a part of the image based upon input data at least for the identified elements in the respective tile.