Apple's STARFlow takes on diffusion models in high-res image generation

When everyone in generative image tech is running full speed down the diffusion model track, Apple is quietly carving out a different path.

While DALL-E and Midjourney rely heavily on iterative denoising diffusion processes, Apple’s machine learning research team just unveiled STARFlow. Developed in collaboration with researchers from UC Berkeley, Georgia Tech, and flow-model pioneers like Laurent Dinh, the system demonstrates that normalizing flows can actually rival state-of-the-art diffusion models in high-resolution image generation.

Normalizing flows transform simple probability distributions into complex ones. Until now, they have been historically sidelined because scaling them to high-resolution, complex images was a mathematical and computational nightmare. The research team claims STARFlow is the first successful demonstration of normalizing flows operating effectively at this scale and resolution.

Under the hood of the flow-based architecture

Apple pulled off this feat using a couple of architectural tricks.

Instead of crunching raw pixels directly, STARFlow operates within the latent space of pretrained autoencoders. This works with compressed image representations and dramatically cuts down computational overhead.

On top of that, researchers introduced what they call a deep-shallow design. A deep Transformer block captures the heavy lifting of model representations, flanked by several shallow Transformer blocks that keep compute costs light without sacrificing output quality.

Unlike diffusion models that take dozens of iterative steps to scrape away noise, STARFlow retains the mathematical purity of normalizing flows. That means it enables exact maximum likelihood training in continuous spaces without discretization. Across both class-conditional and text-to-image prompts, researchers reported sample quality approaching top-tier diffusion models.

Why it matters

Don't expect an immediate update on your phone just yet. This is foundational research, and there is no public app or beta toggle to try right now.

However, this research lands at a fascinating moment for Apple. Following mounting criticism and modest AI announcements at WWDC, recent reports from Bloomberg highlighted how Siri and Apple Intelligence are struggling to match rivals. While OpenAI and Google dominate headlines with massive cloud-hungry models, Apple is exploring fundamental architectures that could suit its own hardware strengths.

We think this mathematical approach is clever. The exact likelihood training inherent to normalizing flows provides a direct way to understand model uncertainty and gain tighter mathematical control over generated content. For on-device computing and enterprise applications where reliability and efficiency matter, this could eventually offer distinct advantages over brute-force diffusion pipelines.

The catch, of course, is the massive head start diffusion models enjoy in commercial toolchains and community mindshare. Proving parity in a paper is one thing; turning a novel flow architecture into an ultra-fast on-device feature for iPhone and Mac chips is another challenge entirely. But it shows Apple has no intention of simply copying everyone else's playbook.