Fessler, Rongrong Wang, and Qing Qu on conditional diffusion models. Its fourth arXiv version was posted on October 20, 2026; the first version was submitted on March 23, 2026.
The authors propose fine-tuning models with an inter-class likelihood-ratio maximization objective, called MCLR. The stated aim is to strengthen class-specific structure in samples generated from the same noise under different labels, reducing reliance on classifier-free guidance (CFG) during sampling.
The paper reports improved guidance-free conditional generation across multiple models and datasets, and says MCLR narrows the gap with CFG. It also proves, under its sample-adaptively weighted ELBO approximation, that the CFG-guided score is the unique minimizer of its objective. The arXiv record does not establish a production implementation, public code, benchmark figures in the abstract, or deployment results.