3D Visual Asset Stylization Using Gram-Matrix Feature Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for 3D stylization fail to generalize across different design styles and often miss capturing important style features due to the complexity of selecting rendered views, leading to ineffective similarity metric estimation.

Innovation Solution

A method using a loss function based on Gram Matrix pairs to identify distinctive-style features, which involves training a convolutional neural network (CNN) to optimize convolutional filter weights, allowing the network to learn and manipulate geometry to match a target visual style.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If hand-crafted shape descriptors are used to evaluate style similarity, then the evaluation can be performed, but the method does not guarantee generalization across different design styles and finding a weighted balance among descriptors is complex

Engineering Contradiction:
Improvegeneralization across different design stylesVSAvoidcomplexity of finding weighted balance among descriptors
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent replaces hand-crafted shape descriptors with a deep learning-based automatic feature extraction system. A neural network model is trained to automatically learn and extract style-specific features from input images, eliminating the need for manual descriptor selection and weight balancing. This substitution enables the system to generalize across different design styles without the complexity of manually tuning descriptor weights.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The neural network model performs self-learning by automatically adapting to different visual styles through training on style-specific images. The system autonomously identifies and extracts discriminative style features without requiring manual intervention for feature engineering or weight adjustment, enabling it to serve multiple style domains universally.

Inventive Principle:
Principle #25Self-service

2Measurement precision

If deep-learning models are used to jointly infer shape descriptors and weights, then style similarity can be computed, but selecting rendered views a priori is complex and directly impacts the final metric by missing important style features

Engineering Contradiction:
Improvestyle similarity evaluation accuracyVSAvoidcomplexity of selecting rendered views
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the essential style features that are discriminative across different designs, rather than relying on pre-selected rendered views. The neural network automatically identifies and extracts the most relevant style characteristics from the input images, eliminating the need for complex view selection and ensuring that important style features are captured.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system changes the approach from fixed pre-selected views to dynamic feature extraction based on learned parameters. The neural network adapts its feature extraction based on the specific style being analyzed, automatically adjusting which features are most discriminative for that particular style domain.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If existing techniques estimate style similarity using fixed methods, then computation is straightforward, but they fail to extract features that are discriminative for a style

Engineering Contradiction:
Improvecomputation efficiencyVSAvoiddiscriminative feature extraction accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent performs preliminary training of the neural network model on style-specific images before actual style similarity evaluation. This preliminary action enables the model to learn discriminative style features in advance, so that during actual use, the system can efficiently compute style similarity with high accuracy using the pre-learned feature extraction capabilities.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4645245A1Visual style transformation using machine learning
Publication Date: 2025.11.05 ROBLOX CORP
  • EP4645245A1 patent drawingFigure 1
  • EP4645245A1 patent drawingFigure 2
  • EP4645245A1 patent drawingFigure 3A

AI summary

According to one aspect of the present disclosure, a method is provided. The method may include receiving a visual asset. The method may include receiving an indicator of a visual style. The method may include identifying a set of distinctive-style features for the visual style using a loss function calculated based on a first summation of differences among a first set of Gram Matrix pairs and a second summation of differences among a second set of Gram Matrix pairs. The method may include modifying a geometry of the visual asset to match the set of distinctive-style features identified for the visual style. After the modifying, the method may include rendering the visual asset that matches the set of distinctive-style features.