LayoutKAG: Enhancing Layout Generation in Large Language Models Through Knowledge-Augmented Generation

Layout generation serves as a cornerstone of automated graphic design, requiring the precise arrangement of multimodal elements to create visually appealing designs that adhere to specified constraints. Existing approaches frequently lack the flexibility to adapt to diverse user constraints. Al-though recent advancements leveraging large language models (LLMs) have shown promise, challenges remain in generating layouts that balance controllability and explainability. In this work, we propose LayoutKAG, a unified framework that integrates domain-specific knowledge and dynamic exemplar retrieval into LLMs through in-context learning. LayoutKAG leverages structured knowledge graphs to encode design prin-ciples and spatial relationships while incorporating historical design exemplars to enhance contextual understanding. Experimental evaluations on benchmark datasets demonstrate that LayoutKAG significantly improves alignment, coherence, and adaptability, highlighting the potential of knowledge-augmented generation to advance the adaptability and effectiveness of LLMs in layout design tasks.