Normalizing Flows with Iterative Denoising

2026-05-06 · Source: Apple Machine Learning Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Iterative TARFlow (iTARFlow) is a new Normalizing Flow (NF) generative model that achieves competitive performance on image modeling tasks, specifically across ImageNet resolutions of 64, 128, and 256 pixels. This model advances prior efforts like TARFlow by maintaining a fully end-to-end, likelihood-based objective during training, distinguishing it from diffusion models. During sampling, iTARFlow employs an autoregressive generation step followed by an iterative denoising procedure, drawing inspiration from diffusion-style methods. The research also includes an analysis of characteristic artifacts produced by iTARFlow, providing insights for future enhancements to Normalizing Flow generative models. Source code for iTARFlow is publicly available on GitHub.

Key takeaway

For research scientists exploring generative models, iTARFlow presents a strong alternative to diffusion models, particularly for image generation tasks. You should consider evaluating iTARFlow's performance on your specific datasets, especially if an end-to-end likelihood-based objective is a priority. Investigating its characteristic artifacts could also inform novel architectural improvements for Normalizing Flows.

Key insights

iTARFlow advances Normalizing Flows with a likelihood-based objective and diffusion-inspired iterative denoising for image generation.

Principles

• Normalizing Flows are viable alternatives to diffusion models.
• End-to-end likelihood objectives are maintained during training.

Method

iTARFlow trains with an end-to-end likelihood objective, then samples via autoregressive generation followed by iterative denoising inspired by diffusion models.

In practice

• Apply iTARFlow for image generation on ImageNet 64, 128, 256.
• Analyze iTARFlow artifacts to guide model improvements.

Topics

• Normalizing Flows
• iTARFlow
• Generative Models
• Image Modeling
• Iterative Denoising

Code references

• apple/ml-itarflow

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows — Normalizing flows can achieve high-quality autoregressive video generation with end-to-end learning and native likelihood estimation.
• A Visual Guide to DiffusionGemma — DiffusionGemma uses iterative diffusion to generate text, optimizing single-user latency by processing 256 tokens concurrently.
• HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising — Conditioning on same-noise-level context in autoregressive diffusion improves temporal consistency and mitigates error propagation.
• CrossFlow: One-Step Generation Across Latent and Pixel Spaces — CrossFlow directly maps noisy latents to pixel-space images, combining latent efficiency with direct pixel supervision in one model.
• Stabilizing, Scaling & Enhancing MeanFlow for Large-scale Diffusion Distillation — A new framework stabilizes and enhances MeanFlow for large-scale diffusion model distillation, improving performance and generalization.
• One-step Language Modeling via Continuous Denoising — Continuous denoising via flow-based models can surpass discrete diffusion in language model quality and speed.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Apple Machine Learning Research.