Normalizing Flows with Iterative Denoising

Researchers introduced iTARFlow, a normalizing-flow image generator that adds iterative denoising at sampling time.

The model combines autoregressive flow-based generation with end-to-end likelihood training and reports strong ImageNet results at 64, 128, and 256 pixels.

The study also examines generated artifacts to point toward future quality improvements.

Overall, the work shows normalizing flows can still compete with diffusion-based image generators for scalable synthesis.