STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
Apple Machine Learning Research
STARFlow2 proposes a unified multimodal generation approach that connects language-model-style autoregression with normalizing flows to reduce fragmentation across text–image generation methods. STARFlow2 is presented as a version “2” that treats autoregressive normalizing flows as autoregressive Transformers using the same left-to-right structure as LLMs. This shifts generation toward a single, Transformer-aligned framework rather than using discrete tokenization, asymmetric text-vs-image pipelines, or adaptation that can harm pretrained understanding.
Why it matters
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 diffusion-based denoising, or degrade pretrained understanding when adapting vision-language models for generation. We observe that autoregressive normalizing flows are autoregressive Transformers—sharing the same causal mask, KV-cache mechanism, and left-to-right structure as LLMs—making them the most…
