Apple's STARFlow2 makes normalizing flows a rival to diffusion
While most of the industry keeps refining diffusion, Apple's research group is betting on a cleaner architecture for models that understand and generate text and images in one system.
Apple's machine learning research group published STARFlow2, a unified multimodal model built on autoregressive normalizing flows — a direct challenge to the diffusion pipelines that power nearly every image generator today. The paper's starting observation is disarmingly simple: an autoregressive normalizing flow already is an autoregressive Transformer, sharing the same causal mask, the same key-value cache, and the same left-to-right structure that large language models are built on. That matters because current unified models mostly bolt a diffusion image generator onto a causal text model and inherit a structural mismatch — text emerges one token at a time while images require an iterative denoising loop, and discrete-token alternatives sidestep that only by giving up visual fidelity. A flow model handles continuous values natively, so a single causal mechanism can carry both modalities end to end.
The architecture, which the team calls Pretzel, vertically interleaves a frozen pretrained vision-language model with a flow stream through residual skip connections, both operating under the same causal mask. That preserves the base model's multimodal understanding instead of degrading it during training for generation — a common failure mode when adapting vision-language models to draw. It also pays off at inference: because text and visual outputs enter the key-value cache directly without re-encoding, interleaved text-and-image sequences stay cache-friendly, the kind of detail that decides whether a unified model is pleasant to serve or painfully slow. Combined with a deep-shallow flow design and a shared latent space, experiments show strong results across both image generation and multimodal understanding benchmarks.
Why it matters: diffusion is so entrenched — from Midjourney to every frontier lab's image and video systems — that alternatives get dismissed as research curiosities. But diffusion pays a per-sample tax in denoising steps, and a flow-based model that matches quality in a single causal pass would make unified generation structurally simpler and cheaper to run. Apple has been building this line quietly: it released the STARFlow-V video model with open weights in December and presented it at CVPR in April. STARFlow2 extends the program to unified text-image generation, and it reads like groundwork for something shippable, not just a workshop paper.
What to watch: whether Apple publishes weights and code the way it did for STARFlow-V — and whether flows close the gap on long, high-resolution sequences where diffusion's lead is still widest.
If one architecture can truly handle both understanding and generation, how long before flows show up in products instead of just papers? Tell us in the comments.
Sources: Apple Machine Learning Research · arXiv · Hugging Face Papers
