Combining plug-in conditioning with classifier-free guidance

Investigate whether the analytic observation-consistency plug-in correction for conditional score-based diffusion models can be combined effectively with classifier-free guidance to improve conditional sampling.

Background

The paper presents analytic plug-in conditioning and classifier-free guidance (CFG) as complementary mechanisms. The plug-in framework derives an observation-consistency correction from the forward corruption model, whereas CFG combines conditional and unconditional score estimates to strengthen conditioning. The paper does not establish whether these mechanisms can be integrated in a principled way or whether their combination improves sampling performance, leaving this question for future investigation.