The idea in plain language.
Choose a training objective that better aligns score learning with likelihood.
How it works
Standard score-based training combines losses across noise levels. The paper studies how the weighting of those losses relates to negative log-likelihood. An appropriate weighting yields a bound that supports approximate maximum-likelihood training. The experiments compare the resulting likelihoods across datasets, model choices, and noising processes.
What to keep in mind
Likelihood and perceptual sample quality measure different properties. Improving one metric does not establish better performance for every application.
Source: Maximum Likelihood Training of Score-Based Diffusion Models. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.