Diffusion models are becoming better not only at generating small molecules and proteins, but conditionally generating within a given training domain. This study suggests that to get better out-of-distribution generation you can use a âguidance functionâ that steers the sample generation process towards properties you want by using unlabeled data and smoothness constraints. The authors say that their approach doesnât add computational overhead in sampling, is easily implemented, and compatible with a range of flavours of diffusion model.
https://arxiv.org/abs/2407.11942