GAN Generator Learning With Portion Codes for Feature Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative adversarial networks (GANs) struggle to accurately associate specific portions of a latent code with individual features, limiting their ability to control and generate images with desired characteristics accurately.
Innovation Solution
A learning system that divides the latent code into portion codes corresponding to specific features, using mapping networks and image synthesis networks to generate images, and employs a generator learning module to enhance the accuracy of feature control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the generator uses the entire latent code to generate images, then the generation process is simple, but the accuracy of associating specific latent code portions with individual features is poor
Solution Approach 1:
The patent segments the latent code into multiple portion codes, where each portion code corresponds to specific features. The generator is divided into multiple image synthesis networks, each responsible for generating images based on specific portion codes. This segmentation enables precise control over individual features while maintaining a structured and manageable generator architecture.
2Manufacturing precision
If the generator is divided into multiple image synthesis networks using portion codes, then feature control accuracy is improved, but the device complexity increases
Solution Approach 1:
The patent applies local quality by assigning different image synthesis networks to handle different features or portions of the latent code. Each network is specialized for specific features, enabling high-precision control over particular aspects of image generation while keeping the overall system organized and manageable through this localized specialization.
3Measurement precision
If mapping networks are introduced to transform latent codes into portion codes, then the learning accuracy is improved, but the training process becomes more complex
Solution Approach 1:
The mapping networks perform preliminary transformation of the latent code into structured portion codes before the main image generation process. This preliminary organization of features enables more efficient and accurate learning during training, as the network receives pre-processed, well-structured input that facilitates faster convergence and reduces overall training time despite the added transformation step.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Provided is a learning system (1) for executing learning of a generator of a generative adversarial network (GAN) which allows a user to control a plurality of features relating to a generated image. A portion code acquisition module (107) acquires a plurality of portion codes respectively corresponding to the plurality of features based on a latent code for generating the generated image and a plurality of mapping networks respectively corresponding to the plurality of features. A generated image generation module (105) generates the generated image based on image synthesis networks that generates the generated image through use of the plurality of portion codes. A generator learning module (106) executes the learning of the generator including the plurality of mapping networks and the image synthesis networks based on the generated image and a trained discriminator of the GAN.