GAN Image Synthesis via Anchor Vectors and Self-Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image editing applications require constant user input and guidance, resulting in unrealistic and distinguishable synthetic images.
Innovation Solution
A computer-implemented method that receives user selections of multiple image anchors, generates synthetic images by finding a vector in a merged space, and evaluates synthetic detectability to provide realistic images with minimal user input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If constant user input and guidance are required during image creation, then the user can control the image generation process, but the user input burden increases and the images become distinguishable as synthetic
Solution Approach 1:
The system performs self-evaluation of synthetic image detectability and automatically iterates on generation without requiring constant user guidance. The generative adversarial network evaluates its own output against real images and adjusts accordingly, enabling the system to serve itself rather than requiring continuous human intervention to achieve realistic results
Solution Approach 2:
The system incorporates feedback mechanisms where the synthetic detectability evaluation is fed back into the generation process. The GAN compares generated images against real images and uses this feedback to refine subsequent generations, creating a loop that continuously improves realism without requiring user input
2Ease of operation
If minimal user input is required for image generation, then the ease of operation improves, but the image quality and realism deteriorate
Solution Approach 1:
The patent replaces the mechanical process of manual user guidance and iterative editing with an automated generative adversarial network system. The GAN uses neural network computations and adversarial training to substitute for what would traditionally require manual user input, achieving both ease of operation and high image quality simultaneously
Solution Approach 2:
The system changes the parameters of image generation from manual control to automated adversarial training parameters. By training the GAN on large datasets of real images and adjusting the adversarial loss parameters, the system achieves realistic image generation with minimal user input, transforming the generation process from mechanical to computational
3Reliability
If synthetic images are generated with high realism, then the image quality improves, but the complexity of the generation system increases
Solution Approach 1:
The generative adversarial network serves multiple functions within a single unified system: it generates images, evaluates their realism, iterates on improvements, and filters results. This multi-functionality consolidates what would traditionally require multiple separate tools and processes into one integrated system, managing complexity through functional consolidation rather than multiplication
Data Source
AI summary
A method including receiving a user selection of multiple image anchors for images within a canvas, and a query for each image anchor, is provided. The method includes finding a vector for the canvas in a merged space associated with the user selection of multiple image anchors, generating a synthetic image for the canvas based on the vector for the canvas in the merged space and an image from an image database, and evaluating a synthetic detectability based on a resemblance of the synthetic image with a real image. The method also includes providing the synthetic image for the canvas to a user when the synthetic detectability is lower than a pre-selected threshold. A system and a non-transitory, computer readable medium storing instructions to perform the above method are also provided.