Diffusion models can be asked not just to generate samples from a desired distribution, but also to avoid an undesired condition. That condition may be specified by a set of reference samples or by a conditional score paired with an unconditional model. Existing approaches inject repulsive terms into the sampling process to push the model away from the unwanted region.
The new paper argues that these repulsive terms are only approximate and proposes a different formulation. By treating the problem as positive-unlabeled learning, the authors derive a guidance term that is asymptotically exact. The approach is designed to work for both ways of specifying the undesired condition.
The source for this item is the paper's abstract only, so the full derivation and any experimental validation are not covered here. The central claim, as stated, is that the method removes the approximation error of earlier negative-guidance techniques.