Grafted Image Generation Model for Faster Multi-Style Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing image generation models require a large number of training samples and time-consuming processes, particularly in face transformation tasks, which can be inefficient and costly.

Innovation Solution

A method involving the training of separate first and second transformation models, followed by grafting these models to create an image generation model that maintains target features while generating images in a desired style, reducing the complexity and time required for training by using a reconstruction model and a grafted transformation model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single transformation model is trained to generate images of different styles, then the model can perform style transformation tasks, but the training process requires a large number of training samples and is time-consuming

Engineering Contradiction:

Improvestyle transformation capabilityVSAvoidtraining time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent divides the transformation model into multiple style-specific sub-models (first transformation model for first style, second transformation model for second style). Each sub-model is trained separately on corresponding style data, avoiding the need to train a single model on all styles simultaneously, thus reducing overall training time while maintaining versatility

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent pre-trains multiple transformation models for different styles before performing style transformation tasks. These pre-trained models are then combined through grafting to create the final image generation model, eliminating the need for time-consuming re-training when new styles are needed

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

Engineering Contradiction:

Improvestyle transformation capabilityVSAvoidtraining sample quantity

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent segments the training data into style-specific subsets and trains separate transformation models for each style. This allows each model to be trained on a smaller, more targeted dataset rather than requiring a single model to process all style data, reducing the total quantity of training samples needed

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent creates multiple transformation models that copy the structure and functionality of a base model but are specialized for different styles. Each model learns style-specific patterns independently, reducing the need for a single model to learn all styles from large quantities of diverse data

Inventive Principle:

Principle #26Copying

3Loss of time

If separate transformation models are trained for different styles and then grafted, then training time and sample requirements are reduced, but the model structure becomes more complex

Engineering Contradiction:

Improvetraining timeVSAvoidmodel structure complexity

Core Design Contradiction:

Loss of timeVSDevice complexity

Solution Approach 1:

The patent merges multiple separate transformation models (first transformation model, second transformation model, etc.) into a unified grafted transformation model. This combining process integrates the strengths of each style-specific model while presenting a cohesive structure that reduces overall complexity compared to managing multiple independent models

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The grafted transformation model serves as a universal model that can generate images in multiple different styles through its integrated architecture. This multi-functional design allows a single model structure to handle various style transformation tasks, reducing the need for separate specialized models and simplifying the overall system

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS12475689B2Method for training image generation model, method for generating images, and devices thereof

Publication Date: 2025.11.18 BIGO TECH PTE LTD

AI summary

Provided is a method for training an image generation model, including: acquiring a first transformation model by training; acquiring a reconstruction model by training based on the first transformation model; acquiring a second transformation model by training; generating a grafted transformation model by grafting the first transformation model with the second transformation model; and generating the image generation model based on the reconstruction model and the grafted transformation model. The first transformation model is configured to generate a first training image according to a first noise sample. The first training image is an image of a first style. The reconstruction model is configured to map an original image sample to a latent variable corresponding to the original image sample. The second transformation model is configured to generate a second training image according to a second noise sample. The second training image is an image of a second style.