Compressed Texture Mipmap Upscaling Without Decompression

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for generating higher resolution images from lower resolution texture data in computer graphics are computationally expensive and require significant storage resources due to the need for decompression and recompression, which is complex and inefficient.

Innovation Solution

Utilizing a neural network to process compressed texture data without decompression or recompression, generating higher resolution textures directly from lower resolution mipmaps, enabling efficient upscaling or downscaling for rendering.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If existing machine learning techniques or other methods are used to produce higher resolution images from lower resolution, then higher resolution textures can be generated, but extra storage space is required for uncompressed versions of both low resolution and high resolution images

Engineering Contradiction:

Improvetexture resolutionVSAvoidstorage space

Core Design Contradiction:

Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent changes the resolution parameter of texture data by processing compressed texture data through a neural network to generate higher resolution versions. The system takes compressed texture data at one resolution and produces compressed texture data at a different resolution without requiring uncompressed storage of intermediate results.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent replaces the traditional mechanical process of decompression-upsampling-recompression with a neural network-based system. The neural network directly processes compressed texture data to generate higher resolution compressed texture data, eliminating the need for uncompressed intermediate storage and the complex decompression-recompression pipeline.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Manufacturing precision

If existing machine learning techniques or other methods are used to produce higher resolution images from lower resolution, then higher resolution textures can be generated, but the process for BCn compressing is quite complex and computationally expensive

Engineering Contradiction:

Improvetexture resolutionVSAvoidcompression process complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent substitutes the complex BCn compression-decompression-upsampling-recompression mechanical process with a neural network system. The neural network is trained to directly generate higher resolution compressed texture data from lower resolution compressed texture data, eliminating the need for the complex traditional pipeline while maintaining computational efficiency.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The neural network is pre-trained on a large dataset of texture pairs at different resolutions. This preliminary training phase allows the network to learn the complex transformations needed for high-quality upscaling, so that during actual use, the complex compression and processing steps are replaced by direct neural network inference, which is computationally more efficient.

Inventive Principle:

Principle #10Preliminary action

3Manufacturing precision

If higher resolution mipmaps are used to render objects near the viewer, then rendering quality is improved, but bandwidth consumption increases

Engineering Contradiction:

Improverendering qualityVSAvoidbandwidth consumption

Core Design Contradiction:

Manufacturing precisionVSLoss of energy

Solution Approach 1:

The patent implements dynamic resolution adjustment by using a neural network to generate higher resolution compressed texture data only when needed for rendering objects near the viewer. The system can dynamically process compressed texture data to the appropriate resolution level based on the rendering requirements, avoiding the need to transmit and store multiple high-resolution versions of all textures.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes the resolution parameter of texture data dynamically based on rendering needs. By processing compressed texture data through a neural network, the system can generate higher resolution versions on-demand, allowing high-quality rendering of nearby objects while maintaining bandwidth efficiency by not transmitting all high-resolution texture data in advance.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260004397A1Machine learning techniques to create higher resolution compressed data structures representing textures from lower resolution compressed data structures

Publication Date: 2026.01.01 SONY INTERACTIVE ENTERTAINMENT LLC

AI summary

Machine learning is used to generate a first mipmap of a texture having a first compression based on a second mipmap of the same texture and having a second compression without using compression or decompression in generating the first mipmap. The first mipmap can then be used to render a computer graphics object.