Parametric Shader Pipeline for Real-Time Frame Transformation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing large neural networks for image transformation in graphics rendering are computationally expensive, undifferentiated, and difficult to modify, making them unsuitable for real-time graphics processing, particularly in resource-constrained settings.

Innovation Solution

A frame transformation pipeline using parametrized shaders, each designed to replicate a specific visual characteristic, is trained using a generative adversarial network, allowing for efficient and modular image enhancement.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If large neural networks are used for image transformation, then accurate replication of target image characteristics is achieved, but computational cost and resource consumption increase significantly

Engineering Contradiction:
Improveaccuracy of image characteristic replicationVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent segments the monolithic large neural network into multiple specialized shaders, each responsible for specific visual characteristics (e.g., lighting, color, texture). This segmentation reduces the computational burden on each individual shader while maintaining overall accuracy through coordinated execution of multiple specialized components.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates a universal shader framework where a small set of parametric shaders can replicate multiple different visual characteristics by adjusting their parameters. This multi-functionality allows the system to achieve accurate image transformation without requiring separate large networks for each characteristic, thereby reducing overall computational cost.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If large neural networks are used for image transformation, then comprehensive learning of all image characteristics is achieved, but the system becomes difficult to modify and adapt

Engineering Contradiction:
Improvecomprehensive learning capabilityVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent divides the comprehensive image transformation task into separate specialized shaders, each handling specific visual characteristics. This segmentation makes the system more adaptable because individual shaders can be modified, added, or removed without affecting the entire system, reducing the complexity associated with modifying large monolithic networks.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic adaptability through parametric shaders that can adjust their behavior based on input conditions and learned parameters. This dynamic approach allows the system to comprehensively learn various image characteristics while maintaining low complexity through the use of simple, configurable shader units rather than fixed complex network architectures.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If large neural networks are used for image transformation, then accurate results are produced, but real-time processing and deployment in resource-constrained settings become infeasible

Engineering Contradiction:
Improveaccuracy of transformation resultsVSAvoidreal-time processing capability
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent replaces expensive large neural networks with cheaper, more efficient parametric shaders that can be rapidly executed. These lightweight shaders maintain acceptable accuracy while enabling real-time processing and deployment on resource-constrained devices such as mobile GPUs and automotive rendering systems.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Solution Approach 2:

The patent achieves accurate image transformation through parameter optimization rather than increasing network size. By adjusting the parameters of simple parametric shaders during training, the system attains accurate results with much lower computational requirements, enabling real-time processing and deployment in resource-constrained environments.

Inventive Principle:
Principle #35Parameter changes

4Adaptability or versatility

If adversarial networks with many parameters are used, then arbitrary characteristics can be learned, but the networks become undifferentiated black boxes that cannot be easily modified

Engineering Contradiction:
Improveability to learn arbitrary characteristicsVSAvoidease of modification
Core Design Contradiction:
Adaptability or versatilityVSEase of operation

Solution Approach 1:

The patent segments the learning process into multiple specialized shaders, each focused on specific visual characteristics. This segmentation creates differentiated, interpretable components that can be individually modified and tuned, eliminating the black box nature of large undifferentiated networks while maintaining the ability to learn arbitrary characteristics through coordinated shader execution.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by designing each shader to handle specific visual characteristics with specialized parameters and algorithms. This localized approach makes each shader interpretable and modifiable for its specific function, while the collection of shaders collectively learns arbitrary characteristics, providing both versatility and ease of operation.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4446996B1Learnable image transformation training methods and systems in graphics rendering
Publication Date: 2026.05.06 IMAGINATION TECH LTD
  • EP4446996B1 patent drawingFigure 1
  • EP4446996B1 patent drawingFigure 2
  • EP4446996B1 patent drawingFigure 3

AI summary

The present disclosure pertains to learnable image transformation training methods and systems in graphics rendering. There is provided a method for training a frame transformation pipeline being part of a graphics processing system and configured to transform rendered frames to produce enhanced frames comprising visual characteristics exhibited in a set of target images. The frame transformation pipeline comprises one or more shaders, defined by a parametrized mathematical function capable of replicating a particular visual characteristic. The training method comprises: receiving input images and target images; applying each shader to the input images to obtain candidate frames, and calculating, at a parametrized discriminator, a similarity indication between characteristics of the candidate frames and the target images. The method further comprises, in dependence on the indication, a parameter update step to parameters of the discriminator and one or more parametrized mathematical functions, wherein the parameter update step is configured to derive parameters the parametrized mathematical function so that the one or more shaders is arranged to impose their respective particular visual characteristics in dependence on an extent to which visual characteristic is exhibited in the target images.