Maximum Likelihood Training of Score-Based Diffusion Models
This work addresses the challenge of improving likelihood estimation in diffusion models for researchers and practitioners in generative modeling, representing an incremental advancement by optimizing an existing training objective.
The paper tackles the problem of training score-based diffusion models by introducing an objective that upper bounds negative log-likelihood, enabling approximate maximum likelihood training. This approach consistently improves likelihoods, achieving negative log-likelihoods of 2.83 and 3.76 bits/dim on CIFAR-10 and ImageNet 32x32, matching state-of-the-art autoregressive models.
Score-based diffusion models synthesize samples by reversing a stochastic process that diffuses data to noise, and are trained by minimizing a weighted combination of score matching losses. The log-likelihood of score-based diffusion models can be tractably computed through a connection to continuous normalizing flows, but log-likelihood is not directly optimized by the weighted combination of score matching losses. We show that for a specific weighting scheme, the objective upper bounds the negative log-likelihood, thus enabling approximate maximum likelihood training of score-based diffusion models. We empirically observe that maximum likelihood training consistently improves the likelihood of score-based diffusion models across multiple datasets, stochastic processes, and model architectures. Our best models achieve negative log-likelihoods of 2.83 and 3.76 bits/dim on CIFAR-10 and ImageNet 32x32 without any data augmentation, on a par with state-of-the-art autoregressive models on these tasks.