STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
Apple's STARFlow2 unifies text and image generation in a single autoregressive model — no more tokenization tradeoffs or diffusion asymmetry.
Why it matters
A research breakthrough in multimodal architecture that eliminates the structural fragmentation plaguing current unified models. If this approach generalizes, it reshapes how vision-language systems handle joint text–image generation at scale.
The key facts
5 to knowApple Research paper on unified multimodal generation
Autoregressive normalizing flows integrated with LLM architecture
Addresses three known problems: discrete tokenization fidelity loss, structural asymmetry (causal + iterative diffusion), vision-language adaptation degradation
Uses same mechanisms as LLMs: causal masking, KV-cache, left-to-right generation
Dated August 25, 2026
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: Unified multimodal models that understand, reason over, and generate interleaved text–image sequences remain structurally fragmented: existing approaches either sacrifice visual fidelity through discrete tokenization, impose structural asymmetry by combining causal text generation with iterative…