MCLR: Improving Conditional Modeling via Inter-Class Likelihood-Ratio Maximization and Unifying Classifier-Free Guidance with Alignment Objectives
Diffusion models often rely on classifier-free guidance (CFG) during inference to achieve strong results. This work introduces MCLR, a training objective that explicitly maximizes inter-class likelihood ratios during training to improve class separation. The research proves that the CFG-guided score is mathematically equivalent to an optimized MCLR objective, suggesting CFG is an implicit alignment procedure.