Autoregressive Image Generation Using Segmented Pixel Groups

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image generation models are computationally intensive due to sequential processing of pixels, leading to high processing times and resource requirements, while sacrificing little quality in output images.

Innovation Solution

The use of autoregressive image generation models that factorize the joint distribution of images into pixel groups, allowing for parallel processing and reducing spatial dependencies, employing multiple sets of convolutional neural networks (CNNs) to upscale images efficiently by merging pixel groups and ordering rules specific to each resolution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If sequential processing of pixels is used in existing image generation models, then processing is straightforward and simple to implement, but processing time increases and resource requirements become high

Engineering Contradiction:
Improveprocessing speedVSAvoidprocessing time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent segments the image generation process into multiple independent processing stages, each handling specific pixel groups rather than processing pixels sequentially. This segmentation allows parallel execution of multiple processing operations simultaneously, dramatically reducing overall processing time while maintaining generation quality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension of parallelism by organizing pixels into groups that can be processed concurrently across multiple processing units. This dimensional transformation from sequential single-pixel processing to parallel group-based processing enables significant speedup without compromising the autoregressive generation quality.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If parallel processing is implemented to reduce processing time, then productivity improves, but spatial dependencies between pixels may be compromised affecting image quality

Engineering Contradiction:
Improveprocessing speedVSAvoidimage quality
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent carefully segments the image into pixel groups that preserve necessary spatial dependencies within each group while allowing independence between groups for parallel processing. This segmentation strategy maintains the autoregressive property within groups (ensuring quality) while enabling parallel execution across groups (improving speed).

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent dynamically adjusts the processing structure by using different ordering rules and processing strategies for different pixel groups based on their spatial relationships. This dynamic approach ensures that spatial dependencies are maintained where critical for image quality while allowing parallelism where independence permits, achieving both speed and quality.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3526765B1Iterative multiscale image generation using neural networks
Publication Date: 2024.04.17 DEEPMIND TECH LTD
  • EP3526765B1 patent drawingFigure 1
  • EP3526765B1 patent drawingFigure 2
  • EP3526765B1 patent drawingFigure 3

AI summary

A method of generating an output image having an output resolution of N pixels x N pixels, each pixel in the output image having a respective color value for each of a plurality of color channels, the method comprising: obtaining a low-resolution version of the output image; and upscaling the low-resolution version of the output image to generate the output image having the output resolution by repeatedly performing the following operations: obtaining a current version of the output image having a current K x K resolution; and processing the current version of the output image using a set of convolutional neural networks that are specific to the current resolution to generate an updated version of the output image having a 2K x 2K resolution.