Conditional Diffusion Models: A Survey of Techniques, Applications, and Challenges
Theodor Panagiotakopoulos, Sotiris Kotsiantis, Αλέξανδρος Γκίλλας, Aris S. Lalos
Abstract
Conditional diffusion models (CDMs) are an emerging family of generative models that enable controllable, data-driven generation across a wide range of modalities. Through conditioning of the generation process on side information such as labels, signals or physical constraints, CDMs promote structured generation, guided sampling and stable inference in settings where complexity, uncertainty and limited data are prevalent. This article provides a complete overview of recent advancements in CDM architectures and, we explore their applications in several areas like computer vision, natural language processing, mechanics, healthcare. Furthermore, we investigate the significant challenges of CDMs. We lastly identify emerging directions and open research problems, with the goal of providing a reference paper for researchers and practitioners engaged in the theory and application of conditional diffusion models.