Autoregressive Generative Modeling with Noise Conditional Maximum Likelihood Estimation

TL;DR

This paper proposes a noise conditional maximum likelihood approach for autoregressive models, enhancing robustness, likelihood, and sample quality through a novel noise-perturbed training scheme and score-based sampling.

Contribution

It introduces a simple modification to MLE by incorporating noise-conditional likelihoods, improving model robustness and sample quality in autoregressive image generation.

Findings

Achieved 3.32 bits per dimension on ImageNet 64x64.

Reduced FID from 37.50 to 12.09 on CIFAR-10.

Models trained with noise conditioning are more robust and generate higher quality images.

Abstract

We introduce a simple modification to the standard maximum likelihood estimation (MLE) framework. Rather than maximizing a single unconditional likelihood of the data under the model, we maximize a family of \textit{noise conditional} likelihoods consisting of the data perturbed by a continuum of noise levels. We find that models trained this way are more robust to noise, obtain higher test likelihoods, and generate higher quality images. They can also be sampled from via a novel score-based sampling scheme which combats the classical \textit{covariate shift} problem that occurs during sample generation in autoregressive models. Applying this augmentation to autoregressive image models, we obtain 3.32 bits per dimension on the ImageNet 64x64 dataset, and substantially improve the quality of generated samples in terms of the Frechet Inception distance (FID) -- from 37.50 to 12.09 on the CIFAR-10 dataset.