STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
STARFlow2 proposes a unified multi-modal generation method aimed at addressing the structural fragmentation issues encountered by existing models in understanding, reasoning, and generating interleaved text-image sequences. This method utilizes autoregressive normalizing flows, which are the most suitable architecture choice due to their shared causal mask, KV cache mechanism, and从左到右 structure.