Generative convolution layer for image generation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative Adversarial Networks (GANs) are powerful tools for image generation but often face challenges in performance and quality.
• Standard convolution methods in GANs can be a bottleneck for generating high-fidelity images tailored to specific latent inputs.

Purpose of the Study:

• To introduce a novel convolution method, Generative Convolution (GConv), designed to enhance GAN performance.
• To demonstrate GConv's effectiveness in improving image quality and generation efficiency in GANs.

Main Methods:

• GConv selects compatible kernels based on a given latent vector.
• Selected kernels are linearly combined to create latent-specific kernels.
• Latent-specific kernels generate latent-specific features, guiding the GAN generator for improved image synthesis.

Main Results:

• GConv significantly improves GAN performance with minimal additional hardware cost.
• The method enhances both unconditional and conditional GANs across various datasets (CIFAR-10, CIFAR-100, LSUN-Church, CelebA, tiny-ImageNet).
• Quantitative evaluations show substantial improvements in Frechet Inception Distance (FID) and Inception Score (IS), e.g., tiny-ImageNet FID improved from 35.13 to 29.76.

Conclusions:

• GConv is a simple yet effective method for boosting GAN performance.
• It can be integrated into existing state-of-the-art GAN architectures without modification.
• The proposed approach offers a promising direction for advancing high-quality image generation in GANs.