Transformer Normalizing Flows for Images

TarFlow is a Transformer-based variant of Masked Autoregressive Flows that enables end-to-end training on image pixels.

It alternates the autoregression direction between Transformer blocks applied to image patches.

Training employs Gaussian noise augmentation and a post-training denoising procedure to improve sample quality.

An effective guidance method is used for both class-conditional and unconditional generation.

TarFlow achieves new state-of-the-art results in image likelihood estimation and generates samples with quality comparable to diffusion models.