Image Style Transfer Training Using Rendered Paired Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video interaction applications lack the capability to support personalized image style transfer, with limited style transfer types and inadequate model training effects.
Innovation Solution
A method and apparatus that utilize simulation models to render target objects in different states and image styles, generating paired sample data for training a machine learning model to ensure high-quality image generation and personalized style transfer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing video interaction applications use standard image style transfer methods, then the implementation is simple, but the image style transfer types are limited and cannot satisfy personalized user requirements
Solution Approach 1:
The patent applies preliminary action by pre-training a machine learning model using paired sample images (first sample images and second sample images) before actual image style transfer operations. This pre-training phase prepares the model to handle multiple personalized style transfer types, enabling the system to adapt to different user requirements without complex real-time adjustments.
Solution Approach 2:
The patent utilizes parameter changes by modifying rendering parameters to generate first sample images and second sample images with different image styles from the same target object. By changing parameters such as lighting, texture, and color during the rendering process, the system creates diverse training data that enables the model to learn multiple style transfer transformations.
2Adaptability or versatility
If more personalized image style transfer types are added, then user requirements are better satisfied, but the model training effect becomes harder to ensure
Solution Approach 1:
The patent applies copying by generating first sample images and second sample images through rendering simulation models of the target object. Instead of requiring actual photographed images in multiple styles, the system creates synthetic copies with controlled variations, ensuring consistent quality and accurate correspondence between paired images for reliable model training.
Solution Approach 2:
The patent implements universality by designing a machine learning model that can handle multiple image style transfer types through a unified training framework. The model is trained on diverse paired sample images representing different styles, enabling it to perform various style transfers (e.g., cartoon, photo, painting styles) using the same trained model, thus ensuring reliability across different personalized requirements.
3Manufacturing precision
If simulation models are rendered in different states and styles to create training data, then sample data quality improves, but the data processing complexity increases
Solution Approach 1:
The patent applies merging by combining the generation of first sample images and second sample images into a unified data processing workflow. Both types of sample images are generated from the same simulation models of the target object in different states, allowing the system to process and pair them systematically, thereby managing complexity while maintaining high data quality.
Data Source
AI summary
The embodiments of the present disclosure relate to an image generation method, apparatus, device, and medium. The method includes: obtaining an initial image of a target object, where the initial image is an image in a first image style; inputting the initial image into a first machine learning model; and obtaining a target image of the target object based on an output result of the first machine learning model, where the target image is an image in a second image style. The first machine learning model is obtained based on a second machine learning model obtained based on training first sample images of the target object and second sample images of the target object. The first sample image and the second sample image are respectively images obtained by rendering simulation models of the target object in different states based on rendering parameters in the first and second image styles.