AI BriefingNormalizing Flows Combined with Iterative Denoising
Key point
Apple researchers proposed iTARFlow, showing competitive results on ImageNet 64, 128, and 256.
Details
As Normalizing Flows draw renewed attention, iTARFlow, a successor to TARFlow, has been proposed.
Training maintains a fully end-to-end likelihood-based objective, while sampling adds iterative denoising after autoregressive generation. Unlike conventional diffusion models, this approach adds refinement capability at the generation stage without changing the training objective.
- Extensive experiments at ImageNet 64/128/256 resolutions confirmed competitive performance.
- The team also analyzed characteristic distortions appearing in generated results, pointing out room for future improvement.
The code has been released at apple/ml-itarflow on GitHub.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Apple Releases LensVLM-9B for Processing Compressed Text Images