Learning priority-aware controllable poster layout generation

Automated graphic layout generation is vital for scalable and personalized multimedia design. Existing approaches often overlook intra-element relationships within layouts and inter-modal dependencies between layout components and visual content. To this end, we introduce a novel priority-aware coarse-to-fine framework that enables both automated and controllable layout synthesis. Our method utilizes an Optimal Transport matcher to align layout elements with corresponding image regions according to their inferred priorities–yielding a structurally coherent yet coarse initial arrangement. This preliminary layout then serves as a strong structural prior for a flow-based generator that refines the composition into an aesthetically pleasing final design. To enable this process automatically, we introduce a Dual-Path Ranker that leverages large language models to assess textual element importance while employing vision models to detect salient visual regions. Extensive experiments on the CGL and PKU poster datasets demonstrate that our approach not only produces high-quality layouts but also provides enhanced adaptability and personalization compared to previous methods.