APPLE STARFLOW · NORMALIZING-FLOW GENERATION
STARFlow: Apple’s normalizing-flow alternative to diffusion
Diffusion dominates image generation, but Apple’s STARFlow line shows that autoregressive normalizing flows (TARFlow) can scale to high-resolution images, then video, and finally to a single model that both understands and generates images and text. Flows give exact likelihoods and cache-friendly autoregressive sampling.
STARFlow vs diffusion and flow matching
Diffusion and flow-matching models learn to denoise or transport noise to data over many steps. A normalizing flow is an invertible network trained by exact maximum likelihood; sampling is a single deterministic pass through the inverse. STARFlow makes this scale by working in a latent space and stacking autoregressive transformer flow blocks. The trade-off is model design complexity rather than sampling steps.
STARFlow2 and unified multimodal models
STARFlow2 interleaves a frozen pretrained VLM stream with a TARFlow stream through residual skip connections under one causal mask. The language side keeps its understanding ability while the flow side generates images, and because both share the KV cache, interleaved text and image generation does not need to re-encode earlier outputs.
Should you deploy it on-device?
Not yet. These are research-scale generators; treat this page as a reference for the papers, code and terminology rather than a deployment guide. For on-device vision that runs today, use FastVLM, MobileCLIP2 or Depth Pro.
Questions
Is there a STARFlow demo?
Not on this site. Image generation at this scale needs server GPUs, which conflicts with FastVLM.net’s on-device focus. Follow the GitHub repository for official samples and checkpoints.