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.
- MCLR fine-tunes diffusion models to maximize inter-class likelihood ratios, internalizing classifier-free guidance effects.
- Paper proves CFG-guided score is the exact optimum of a weighted MCLR objective, unifying training and inference-time alignment.
- On EDM2-L ImageNet-512, MCLR hits 42.50 FD-DINOv2 unguided vs CFG's 39.86 and base 67.70.
- On VAR-d24 and SiT-XL/2+REPA, MCLR actually beats CFG on best-case FD-DINOv2 scores.
- Reveals previously unknown equivalence between CC-DPO and Conditional Contrastive Alignment at the population level.
- Limitation: produces a fixed model without CFG's inference-time fidelity-diversity knob, and doubles training batch cost.
Researchers from the University of Michigan and Michigan State University propose a fine-tuning method that teaches conditional generative models to separate classes more sharply. The MCLR paper connects this training objective to classifier-free guidance, or CFG, and reports comparable or better unguided results on several ImageNet benchmarks. Removing CFG can cut the model evaluations required at each diffusion sampling step.
Why class labels fade
Diffusion models learn to reverse a process that gradually adds noise to data. At the theoretical optimum, a conditional score network contains enough information to sample from p(x|c), the data distribution for condition c. Practical models approximate that score, and their unguided samples often fail to express the condition strongly enough.
CFG sharpens conditional generation by combining conditional and unconditional scores at every reverse step. A common parameterization is:
s_cfg = s_uncond + g * (s_cond - s_uncond)Increasing the guidance scale g generally improves condition adherence and fidelity while reducing diversity. Prior work cited by the authors reports FID reductions of up to 75%. The method requires both conditional and unconditional model evaluations at each step, roughly doubling inference arithmetic even when implementations batch the two evaluations. Its modified score also lacks a standard corresponding forward diffusion process, which complicates its probabilistic interpretation.
The paper attributes CFG’s effectiveness to weak separation among learned class-conditional distributions. Samples generated from the same initial noise under different ImageNet labels can retain similar global layouts while losing features that distinguish the requested classes. MCLR directly penalizes that overlap during fine-tuning.
A likelihood ratio in the loss
Maximum Inter-Class Likelihood Ratio, abbreviated MCLR, trains the model to assign each sample a higher conditional likelihood under its correct class than under a mismatched class. For a sample
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
