A Systematic Comparison and Fusion of Generative Adversarial Networks and Diffusion Models for Image Generation

Abstract

Generative Adversarial Networks (GANs) and diffusion models are two mainstream approaches in image generation. Although their underlying principles differ, both can be unified under the framework of particle models (PMs). GANs enable fast generation but suffer from unstable training and mode collapse, whereas diffusion models produce high-quality, diverse samples at the cost of slow sampling. Although various diffusion model variants have been proposed to accelerate sampling, a systematic comparison under identical experimental conditions remains lacking. This paper presents a fair, comprehensive comparison of GAN, Denoising Diffusion Probabilistic Model (DDPM), Denoising Diffusion Implicit Model (DDIM), and a proposed fusion model on the Fashion-MNIST dataset, evaluating generation quality with Fréchet Inception Distance (Feature-FID) and Kernel Inception Distance (Feature-KID) and analyzing the effects of sampling steps and stochasticity. Results show that DDIM with stochastic sampling (η=1) achieves the lowest KID of 575.51±5.69, outperforming deterministic DDIM and the standard GAN. The proposed fusion model, combining GAN warm-up with DDIM refinement, reduces the number of sampling steps while achieving a KID of 672.17±65.15, offering a trade-off between generation quality and inference efficiency.