Apple researchers have introduced STARFlow2, an experimental architecture that handles text and continuous image generation with a shared left-to-right mechanism. The design aims to avoid the split structure commonly used when a language model generates text but a separate diffusion process produces images.
STARFlow2 combines a frozen pretrained vision-language stream with an autoregressive normalizing-flow stream through residual connections. Both operate under the same causal mask, so text and visual outputs can enter the key-value cache directly without being encoded again. A normalizing flow is a reversible generative model that transforms a simple distribution into complex data such as images.
Keeping the vision-language model frozen is intended to preserve its existing understanding while the additional stream supplies higher-fidelity continuous image generation. The researchers report results across image-generation and multimodal-understanding benchmarks and argue that autoregressive flows can support interleaved sequences containing both words and pictures.
This is a research system, not an announced Apple product. The published description establishes architectural and benchmark results, but it does not provide consumer availability, deployment cost or evidence that the approach outperforms production multimodal systems in every setting.