GANs have inherent drawbacks such as training instability and susceptibility to mode collapse, which prompted the research community to shift toward diffusion models that generate images through a gradual denoising Markov chain process.