TokenFlow bridges multimodal understanding and generation using an innovative dual-codebook architecture that decouples semantic and pixel-level feature learning while maintaining alignment through a shared mapping mechanism. It surpasses flagship models like LLaVA-1.5 and EMU3 on visual comprehension tasks and achieves performance comparable to SDXL on text-to-image generation at 256×256 resolution.