Policy-guided diffusion combines elements of diffusion models and reinforcement learning to control the sample generation process in deep generative models. Diffusion models work by gradually converting random noise into a structured output through a series of learned reverse diffusion steps. In policy-guided diffusion, a policy dictates certain aspects of these steps based on specific criteria or objectives, which could include maximizing the likelihood of generating samples with certain features or adhering to particular constraints. This approach is particularly useful in scenarios where the generative process must satisfy predefined conditions, such as in controlled image generation or in generating synthetic data that adheres to complex rules.
While diffusion models have been a subject of research since the early 2010s, the concept of integrating a guiding policy became more pronounced with the advent of more sophisticated reinforcement learning techniques in the late 2010s. The specific term "policy-guided diffusion" and its applications have become more common in the academic literature and practical applications over the past few years.
The development of policy-guided diffusion involves interdisciplinary contributions from the fields of generative models and reinforcement learning. Specific key contributors often include researchers from both academic institutions and technology companies who specialize in deep learning, generative modeling, and AI-driven optimization. The exact originators of the term may not be well-defined as it represents a convergence of ideas rather than a single invention.