Apple researchers introduce iTARFlow, advancing Normalizing Flows with iterative denoising for image generation

Apple researchers presented iTARFlow, an iterative variant of TARFlow (Transformer Autoregressive Flow) that combines a likelihood-based training objective with iterative denoising during sampling, inspired by diffusion-style methods.

Unlike diffusion models, iTARFlow maintains a fully end-to-end, likelihood-based objective during training while performing autoregressive generation followed by iterative denoising during the sampling phase.

iTARFlow achieves competitive performance across ImageNet resolutions of 64, 128, and 256 pixels, demonstrating Normalizing Flows as viable alternatives to other generative methods. Code is available at https://github.com/apple/ml-itarflow.