Diffusion models operate on a theoretical framework where image generation is modeled as reversing Brownian motion within high-dimensional space by predicting and removing noise added via stochastic differential equations (SDEs). The core mechanism relies on learning a time-varying vector field, specifically the score function of the data distribution at each timestep $t$, which allows for deterministic sampling through ordinary differential equation-based flows like DDIM. This domain belongs to probabilistic generative modeling, bridging statistical mechanics and deep learning by utilizing flow matching to steer latent trajectories from an isotropic Gaussian prior toward complex image manifolds while decoupling global structure recovery from class-specific conditioning via classifier-free guidance.
Diffusion models operate on a theoretical framework where image generation is modeled as reversing Brownian motion within high-dimensional space by predicting and removing noise added via stochastic …