Hermes Flow is a universal multimodal large language model alignment framework capable of autonomously generating homologous preference data. Through self-play iterative optimization and paired DPO techniques, it seamlessly bridges the gap between multimodal understanding and generation.