The underlying architecture of GPT-Image-2 integrates the semantic understanding capabilities of large language models, enabling it to automatically convert natural language descriptions into visual generation instructions.