基于历史条件多模态大模型的非马尔可夫多轮对话图像生成

Multi-round conversational image generation necessitates models to adhere to user instructions across multiple interactions, grounding their outputs in interleaved text and images that accumulate as chat history. While contemporary multimodal large language models (MLLMs) are capable of generating and editing images, most existing multi-turn benchmarks and training paradigms are effectively Markovian. This implies that the subsequent output primarily hinges on the most recent image, fostering shortcut solutions that neglect the comprehensive conversational history. When the full dialogue history is disregarded, models struggle to capture long-range dependencies and intricate contextual shifts, significantly compromising the semantic consistency and coherence of generated images. For instance, in scenarios where users request multiple fine-grained modifications to the same image, if the model exclusively focuses on the last modification instruction and the preceding image, overlooking all prior modification history, the generated result may deviate from the user's initial intent or even exhibit semantic conflicts. This limitation of the Markovian assumption prevents models from truly understanding and leveraging the cumulative context of the conversation, thereby restricting their performance in complex, long-range dialogue settings. To surmount this challenge, new model architectures and training strategies are required that enable models to effectively integrate and utilize complete conversational history information for more accurate and coherent multi-round conversational image generation. This includes designing mechanisms capable of encoding and remembering multi-modal historical data, as well as establishing metrics to evaluate long-term consistency in non-Markovian dialogues.