Stylized Training Data Synthesis for Disentangled Style Learning

Solution Overview

Problem

Conventional techniques for searching and identifying digital content based on style suffer from inaccuracies due to the similarity of subject matter, leading to inefficient use of computational resources, increased power consumption, and user frustration.

Innovation Solution

The development of stylized training data synthesis techniques, which involve selecting a style example and subject matter examples, synthesizing stylized training data using neural transfer techniques, and training a machine-learning model to learn a representation of style that is disentangled from subject matter.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional search techniques are used to identify digital content based on style, then the search functionality is provided, but inaccuracies occur due to subject matter similarity leading to inefficient computational resource use and increased power consumption

Engineering Contradiction:

Improvestyle identification accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent segments the style identification task into two independent components: subject matter identification and style identification. By training separate machine learning models for each component and processing them in parallel, the system avoids redundant computational operations on identical content regions, thereby reducing power consumption while improving style identification accuracy through specialized processing.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by pre-training dedicated machine learning models for subject matter identification and style identification before actual search operations. This pre-training phase creates optimized feature extractors that quickly process content during runtime, reducing the computational energy required during actual searches while maintaining high accuracy in distinguishing style from subject matter.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If conventional search techniques are used to identify digital content based on style, then the search functionality is provided, but computational resources are used inefficiently

Engineering Contradiction:

Improvestyle identification accuracyVSAvoidcomputational efficiency

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the style identification task into two independent components: subject matter identification and style identification. By training separate machine learning models for each component and processing them in parallel, the system avoids redundant computational operations on identical content regions, thereby reducing power consumption while improving style identification accuracy through specialized processing.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by pre-training dedicated machine learning models for subject matter identification and style identification before actual search operations. This pre-training phase creates optimized feature extractors that quickly process content during runtime, reducing the computational energy required during actual searches while maintaining high accuracy in distinguishing style from subject matter.

Inventive Principle:

Principle #10Preliminary action

3Measurement precision

If conventional search techniques are used to identify digital content based on style, then the search functionality is provided, but user frustration increases due to inaccuracies

Engineering Contradiction:

Improvestyle identification accuracyVSAvoiduser efficiency

Core Design Contradiction:

Measurement precisionVSEase of operation

Solution Approach 1:

The patent segments the style identification task into two independent components: subject matter identification and style identification. By training separate machine learning models for each component and processing them in parallel, the system avoids redundant computational operations on identical content regions, thereby reducing power consumption while improving style identification accuracy through specialized processing.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by pre-training dedicated machine learning models for subject matter identification and style identification before actual search operations. This pre-training phase creates optimized feature extractors that quickly process content during runtime, reducing the computational energy required during actual searches while maintaining high accuracy in distinguishing style from subject matter.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20250182461A1Stylized training data synthesis for training a machine-learning model

Publication Date: 2025.06.05 ADOBE INC

AI summary

Stylized training data synthesis techniques are described for training a machine-learning model. In one or more examples, a style training system selects a style example exhibiting a style that is to be subject of training a machine-learning model. The style training system also selects a collection of subject matter examples having different instances of subject matter. The style training system then synthesizes stylized training data based on the style example and subject matter examples, e.g., using a neural transfer technique. The stylized training data is then usable to train a machine-learning model to learn a representation of style that has limited influence by the subject matter being expressed by the digital content.