Reliable out-of-distribution sampling in guided diffusion

Establish sampling procedures for denoising diffusion models that reliably generate high-value samples from regions beyond the training data distribution, enabling robust conditional generation in out-of-distribution settings rather than relying solely on modifications to the diffusion process itself.

Background

Diffusion models are effective at modeling complex data distributions and have demonstrated strong performance in unconditional and conditional generation tasks. Conditional generation typically relies on guidance functions learned from labeled data to steer sampling toward desirable properties.

In many practical domains, such as molecular and protein design, labeled data is scarce and biased, making reliable exploration of high-value regions outside the training distribution difficult. The paper identifies this as an open challenge and proposes context-guided diffusion (CGD) to improve out-of-distribution generalization by regularizing guidance models with unlabeled context data and smoothness constraints.

The explicit open challenge highlights the broader need for principled methods that ensure reliable conditional sampling into high-value, out-of-distribution regions, beyond approaches that primarily modify the diffusion process itself.