Virtual Image Generation by Fusing Body Features and Target Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to efficiently generate high-quality virtual images retaining object body features and specific attributes due to the difficulty in obtaining sufficient sample images with these characteristics, leading to lower efficiency and quality in image generation models.
Innovation Solution
Fusing a first image generation model trained on original sample images for object body features with a second model trained on virtual sample images for target attributes to create a target image generation model that focuses on both aspects, thereby improving efficiency and quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a large number of sample images retaining both object body features and target attributes are used for training, then the quality of generated virtual images is improved, but the time and resources required to obtain and process these sample images increases significantly
Solution Approach 1:
The patent divides the training process into two separate stages: first training an image generation model to preserve object body features using original sample images, then training a second model to add target attributes using virtual sample images. This segmentation allows each model to specialize in one aspect, achieving high-quality results without requiring a single model to process all complex requirements simultaneously, thus reducing the time and resources needed.
Solution Approach 2:
The patent performs preliminary action by first generating virtual sample images with target attributes through the first image generation model before using them to train the second model. This preliminary generation of training data enables the second model to focus specifically on attribute addition without needing access to rare real-world samples that simultaneously contain both object body features and target attributes, significantly reducing data collection time.
2Adaptability or versatility
If a single image generation model is trained to simultaneously preserve object body features and add target attributes, then the model complexity increases, but the training efficiency and convergence speed decrease
Solution Approach 1:
The patent segments the image generation task into two specialized models: the first model handles object body feature preservation, while the second model handles target attribute addition. This division allows each model to be simpler and more efficient at its specific task, improving training efficiency and convergence speed while maintaining overall system versatility through model composition.
Solution Approach 2:
The patent combines two specialized image generation models into a unified training pipeline where the first model's output serves as input for the second model. This merging approach achieves the versatility of handling both object body features and target attributes while maintaining the training efficiency of specialized models, as each model trains independently on its specific task rather than competing for the same training resources.
3Manufacturing precision
If original sample images with both object body features and target attributes are collected directly, then the quality requirement is met, but the difficulty and cost of obtaining sufficient training data increases
Solution Approach 1:
The patent performs preliminary action by using the first image generation model to synthesize virtual sample images with target attributes before they are needed for the second stage training. This preliminary generation creates abundant training data that would be difficult or impossible to obtain from real-world sources, eliminating the need to search for and collect rare samples that simultaneously contain both object body features and specific target attributes.
Solution Approach 2:
The patent uses the first image generation model to create copies of object body features in virtual sample images, which then serve as training data for the second model. Instead of requiring original real-world samples with target attributes, the system creates synthetic copies that preserve the essential characteristics needed for training, significantly reducing data acquisition difficulty and cost while maintaining image quality.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The present application relates to the technical field of artificial intelligence, and discloses a method and apparatus for obtaining a virtual image, a computer device, a computer-readable storage medium, and a computer program product. The method comprises: obtaining a target image generation model, the target image generation model being obtained by fusing a first image generation model and a second image generation model, the first image generation model being obtained by training on the basis of sample original images, the sample original images retaining object body features of sample objects, and the second image generation model being obtained by training on the basis of sample virtual images having target attributes; and obtaining, on the basis of the target image generation model, a target virtual image corresponding to an original image of a target object, the target virtual image retaining an object body feature of the target object and having a target attribute.