Diffusion Models for High-Fidelity Image Synthesis: From Score Matching to Latent Generative Modeling

Abstract

Deep generative modeling underwent a substantial transition by 2022 as diffusion-based approaches emerged as strong alternatives to generative adversarial networks, variational autoencoders, and autoregressive image models. Diffusion models formulate generation as the reversal of a gradual noise-corruption process, allowing complex data distributions to be learned through repeated denoising operations. This review examines the development of diffusion-based generative modeling through 2022, beginning with nonequilibrium thermodynamic formulations and denoising score matching and progressing toward denoising diffusion probabilistic models, score-based stochastic differential equations, deterministic implicit sampling, classifier guidance, and latent diffusion. Denoising diffusion probabilistic models demonstrated high-quality synthesis using parameterized reverse diffusion, while improved parameterizations and classifier guidance substantially enhanced sample fidelity. Score-based generative modeling unified several related approaches through stochastic differential equations. Denoising diffusion implicit models reduced sampling requirements by defining non-Markovian reverse processes. Subsequent classifier-free conditioning enabled controllable generation without a separate classifier. Latent diffusion further addressed the substantial computational cost of pixel-space models by performing the generative process within compressed perceptual representations. Concurrent text-conditioned systems illustrated the capacity of diffusion architectures to connect natural-language representations with high-resolution image synthesis. The review compares these methods in terms of likelihood formulation, sampling efficiency, conditioning mechanisms, image fidelity, latent representation, and computational requirements. Challenges remaining in 2022 included slow iterative inference, expensive large-scale training, evaluation of semantic alignment, dataset bias, memorization, and the computational accessibility of high-resolution generation.