STARFlow2: Bridging language models with normalizing flows for unified multimodal generation
STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
AI BriefingSTARFlow2 proposes unifying autoregressive normalizing flows with language models for multimodal generation. The approach observes that autoregressive normalizing flows and LLMs share causal masks, KV-cache mechanisms, and a left-to-right structure, bridging the architectural gap between text and image generation while avoiding the loss of visual fidelity caused by discrete tokenization.
Source: Apple Machine Learning Research · machinelearning.apple.com