ICG: Improving Cover Image Generation via MLLM-based Prompting and Personalized Preference Alignment

A novel framework named ICG has been introduced to enhance personalized cover image generation by integrating MLLM-based prompting with user preference alignment. This approach utilizes semantic features from item titles and reference images, refining them with user embeddings to produce contextually relevant covers. The framework also employs a multi-reward learning strategy to overcome the challenges posed by the lack of labeled supervision.

The development of ICG is significant as it addresses the underexplored area of personalized cover image generation, which is crucial for increasing user engagement on digital platforms. By leveraging advanced AI techniques, ICG aims to improve the quality and relevance of generated images, potentially transforming how content is presented and consumed online.

This advancement reflects broader trends in AI, particularly in the realm of multimodal large language models and diffusion models, which are increasingly being utilized to enhance content creation. The integration of user preferences into AI-generated outputs raises important discussions about personalization in digital media, as well as the ongoing challenges of ensuring alignment and safety in AI systems.