GAN Compression With Balanced Generator-Discriminator Pruning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Generative Adversarial Networks (GANs) suffer from high computational cost and bulky memory usage, leading to performance degradation and mode collapse when compressed on devices with limited computing resources, such as mobile phones, due to the imbalance in generator and discriminator compression.

Innovation Solution

Perform pruning on the generator and configure the states of convolution kernels in the discriminator to maintain Nash equilibrium, ensuring the generator and discriminator maintain a balanced performance by adjusting retention factors and weight parameters to minimize loss differences.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If a large-scale network model is used to improve image generation quality, then the generation quality is improved, but the memory resource consumption increases

Engineering Contradiction:
Improveimage generation qualityVSAvoidmemory resource consumption
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent divides the large-scale network model into multiple smaller network models by splitting the input image into multiple regions. Each smaller network model processes a specific region, reducing the memory footprint of individual models while maintaining the overall quality through coordinated processing of all regions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from a single large-scale model approach to a multi-model architecture that processes different spatial regions independently. This dimensional transformation allows parallel processing across multiple models, reducing peak memory consumption while achieving comparable generation quality through the aggregation of regional results.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Manufacturing precision

If a large-scale network model is used to improve image generation quality, then the generation quality is improved, but the computation resource consumption increases

Engineering Contradiction:
Improveimage generation qualityVSAvoidcomputation resource consumption
Core Design Contradiction:
Manufacturing precisionVSPower

Solution Approach 1:

The patent segments the computation task by dividing it into multiple smaller network models, each handling a specific region. This segmentation reduces the computational burden on any single model, allowing for more efficient resource utilization while maintaining overall generation quality through the combination of regional outputs.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by having each smaller network model focus on processing only its designated region rather than the entire image. This partial processing approach reduces the computational load per model while the aggregation of all regional processing results achieves the desired overall image generation quality.

Inventive Principle:
Principle #16Partial or excessive action

3Manufacturing precision

If a large-scale network model is used to improve image generation quality, then the generation quality is improved, but the time consumption increases

Engineering Contradiction:
Improveimage generation qualityVSAvoidtime consumption
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent segments the image processing task into multiple parallel regions, each handled by a smaller network model. This segmentation enables parallel computation across multiple models, reducing the overall processing time compared to a single large model processing the entire image sequentially, while maintaining quality through coordinated regional processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces parallelism as a new dimension by deploying multiple smaller network models simultaneously to process different regions. This dimensional change from sequential single-model processing to parallel multi-model processing significantly reduces time consumption while achieving comparable generation quality through the integration of regional results.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4339836B1Network model compression method, apparatus and device, image generation method, and medium
Publication Date: 2026.05.13 BEIJING ZITIAO NETWORK TECH CO LTD
  • EP4339836B1 patent drawingFigure 1
  • EP4339836B1 patent drawingFigure 2
  • EP4339836B1 patent drawingFigure 3

AI summary

A network model compression method, apparatus and device, an image generation method, and a medium. The network model compression method comprises: performing pruning processing on a first generator, so as to obtain a second generator (S 1 10); and configuring the states of convolution kernels in a first discriminator, such that some of the convolution kernels are in an activated state and the other convolution kernels are in an inhibited state, so as to obtain a second discriminator (S120), wherein a loss difference between the first generator and the first discriminator is a first loss difference, a loss difference between the second generator and the second discriminator is a second loss difference, and the absolute value of the difference between the first loss difference and the second loss difference is less than a first preset threshold value. By means of the method, a generator and a discriminator can be collaboratively compressed, and therefore the compressed generator and the compressed discriminator can maintain a Nash equilibrium, thereby avoiding the phenomenon of mode collapse.