STARFlow2 is a research framework for unified text-image generation using autoregressive normalizing flows. It reframes normalizing flows as autoregressive Transformers with shared causal masking and KV-cache, avoiding the visual fidelity loss of discrete tokenization and the asymmetry of diffusion-based methods. This points toward structurally unified multimodal models that preserve pretrained understanding while generating interleaved content.
Opening Kapyn…