STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation

Apple researchers have introduced STARFlow2, a framework designed to integrate language models with normalizing flows to create a unified system for processing interleaved text and image data. This approach aims to overcome current architectural limitations where models typically struggle to balance high-quality image generation with fluid text reasoning. By moving away from discrete tokenization, the method seeks to improve visual fidelity and maintain structural consistency in multimodal generation.