Self-Converging Generative Network for Sharp Image Synthesis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current generative models, such as VAE and GAN, face issues like blurry images due to direct pixel-wise learning and mode collapsing, which are difficult to resolve with existing training methods, especially in balancing generator and discriminator parameters.

Innovation Solution

A self-converging generative network (SCGN) is introduced, where a latent space and image space converge using a single network, with a loss function derived from pixel-wise loss and Kullback-Leibler divergence, allowing for one-to-one mapping and training without sampling, thus alleviating blurriness and mode collapsing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If VAE uses pixel-wise loss for direct image learning, then training is simplified, but generated images become blurry

Engineering Contradiction:
Improvetraining simplicityVSAvoidimage sharpness
Core Design Contradiction:
Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent merges the generator and discriminator networks into a single unified network structure. The generator portion learns to create images while the discriminator portion provides adversarial feedback within the same network, allowing pixel-wise learning without sampling while maintaining sharpness through internal adversarial mechanisms.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The network dynamically switches between generation mode and discrimination mode during training. The same network parameters are used for both functions, with dynamic adjustment of which component is active based on the training phase, enabling both sharp image generation and simplified training.

Inventive Principle:
Principle #15Dynamics

2Manufacturing precision

If GAN uses adversarial loss for indirect image learning, then image sharpness is improved, but training stability deteriorates due to generator-discriminator balancing difficulties

Engineering Contradiction:
Improveimage sharpnessVSAvoidtraining stability
Core Design Contradiction:
Manufacturing precisionVSStability of the object's composition

Solution Approach 1:

By merging the generator and discriminator into a single network, the patent eliminates the parameter mismatch problem between separate networks. The unified structure ensures consistent parameter updates and eliminates the need for careful balancing of two separate training processes, thereby improving training stability while maintaining image sharpness.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The single network performs both generation and discrimination functions universally. The same network parameters serve dual purposes: generating images when in generation mode and evaluating authenticity when in discrimination mode, eliminating the need for separate network tuning and improving overall training stability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Manufacturing precision

If GAN structure is used for image generation, then realistic images are produced, but mode collapsing occurs frequently

Engineering Contradiction:
Improveimage realismVSAvoidmode collapsing resistance
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The unified network structure prevents mode collapsing by ensuring that the generator and discriminator share the same parameter space. This allows the network to explore the full data manifold more effectively during training, maintaining diversity in generated samples while producing realistic images, thereby eliminating mode collapsing.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The integrated discriminator provides continuous feedback within the same network, allowing the generator to adjust its output distribution to cover all modes of the data. This internal feedback mechanism ensures comprehensive coverage of the data manifold, preventing mode collapsing while maintaining image realism.

Inventive Principle:
Principle #23Feedback

4Adaptability or versatility

If latent space sampling is performed in VAE, then multiple images can be learned in single latent space, but image sharpness deteriorates

Engineering Contradiction:
Improvelatent space utilizationVSAvoidimage sharpness
Core Design Contradiction:
Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The network dynamically determines whether to sample from the latent space or use direct pixel-wise learning based on the training phase and mode. During adversarial training, the network uses direct learning without sampling to maintain sharpness, while still benefiting from latent space representation when needed for generation versatility.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentUS20230297829A1Method and system for training self-converging generative network
Publication Date: 2023.09.21 HANSUNG UNIV IND UNIV COOPERATION FOUND
  • US20230297829A1 patent drawing
  • US20230297829A1 patent drawing
  • US20230297829A1 patent drawing

AI summary

A method and system for training a self-converging generative network are disclosed. A method for training a self-converging generative network according to one embodiment of the present invention comprises the steps of: pairwise-mapping training images and latent space vectors constituting a training data set; defining a loss function for a generator of the self-converging generative network and a loss function for a latent space; and training weights and latent vectors of the self-converging generative network using the loss function for the generator and the loss function for the latent space.