Michigan Researchers' MCLR Bakes Guidance Into Training, Cutting Diffusion Inference Costs

A new training objective called MCLR reproduces classifier-free guidance effects at training time, revealing CFG as implicit contrastive alignment.