Customized Generative Models for Individual Image Attributes

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current generative models lack the ability to create customized images that reflect individual human impressions and perceptions, failing to accurately capture the unique attributes and judgments of specific individuals or groups, which is crucial for personalized product design and marketing.

Innovation Solution

A method is developed to construct customized generative models by training regression models on user-specific data, using synthetic images and latent representations, allowing for the generation of images that embody target attributes and perceptions, such as making a car appear fast or expensive, tailored to individual or group demographics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a generative model is trained on general data to create diverse images, then image diversity is improved, but the ability to capture individual human impressions and perceptions deteriorates

Engineering Contradiction:

Improveimage diversityVSAvoidaccuracy of capturing individual human impressions

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent segments the generative model into two distinct components: a base generative model trained on general data for diversity, and a customization layer trained on individual user data for precision. The base model generates diverse images while the customization layer adjusts them to match individual human impressions, resolving the contradiction between diversity and individualization.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The customization layer is nested within the base generative model architecture. The base model provides the foundational image generation capabilities, while the customization layer is embedded within it to add individualized adjustments. This nested structure allows both general diversity and individual precision to coexist within a unified system.

Inventive Principle:

Principle #7Nested doll (Nesting)

2Adaptability or versatility

If a generative model is trained on diverse data to represent multiple attributes, then attribute coverage is improved, but the ability to accurately represent specific individual attributes deteriorates

Engineering Contradiction:

Improveattribute coverageVSAvoidaccuracy of representing specific individual attributes

Core Design Contradiction:

Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The customization layer applies local quality adjustments to specific attributes of generated images based on individual user preferences. While the base model provides broad attribute coverage, the customization layer selectively refines specific attributes (such as speed, cost, or other perceptual qualities) to match individual human impressions, achieving both coverage and precision.

Inventive Principle:

Principle #3Local quality

3Productivity

If synthetic images are generated using standard generative models, then image generation speed is improved, but the ability to embody target attributes and perceptions deteriorates

Engineering Contradiction:

Improveimage generation speedVSAvoidalignment with target attributes and perceptions

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The base generative model performs preliminary image generation quickly using pre-trained knowledge, producing images that broadly represent target attributes. The customization layer then applies targeted adjustments to align with specific individual perceptions. This preliminary action allows fast generation while maintaining precision through subsequent refinement.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system changes parameters in two stages: first generating images with broad attribute parameters using the base model for speed, then adjusting specific parameters (such as perceived speed, cost, or other human impression attributes) using the customization layer. This parameter change approach maintains generation speed while improving attribute alignment precision.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20250104404A1Individualized generative models for image generation and manipulation

Publication Date: 2025.03.27 UNIVERSITY OF CHICAGO

AI summary

Systems and methods are provided herein for training a customized model. A method of constructing a customized generative model, comprising reading a plurality of synthetic images and associated latent representations; presenting each of the plurality of synthetic images to one or more users via a client computing platform; reading a plurality of inputs characterizing a plurality of values for a plurality of associated attributes of each of the plurality of synthetic images; based on the values of the associated attributes and the latent representations, training a regression model to predict the values of the attributes from the latent representations.