TarFlow: Transformer Normalizing Flows for Images

The TarFlow model applies Transformer-based autoregressive flows on image patches with alternating directions to model pixel distributions.

Training enhancements include Gaussian noise augmentation, a post-training denoising step, and a guidance mechanism for better sample quality.

TarFlow sets new state-of-the-art results in image likelihood estimation, surpassing previous normalizing flow methods.

TarFlow alone can generate diverse, high-quality images comparable to diffusion models.