LayoutPrompter: Awaken the Design Ability of Large Language Models

Conditional graphic layout generation, which automatically maps user constraints to high-quality layouts, has attracted much attention in recent years. Despite good performance, recent work still suffers some pivotal challenges. First, the neural models customized for this task require a large amount of layout data for model training, which is time-consuming and expensive. Second, the previous approaches usually do not have strong cross-domain generalization ability (e.g., from UI to Document). In this work, we propose LayoutPrompter to address the aforementioned issues by simply prompting GPT-3 text-davinci-003 model with a few demonstration examples. By meticulously designing the prompting strategy, our approach can generate high-quality, cross-domain graphic layouts without any model training or fine-tuning. Although remarkably simple, the experiments show that LayoutPrompter is competitive with state-of-the-art approaches on five traditional conditional layout generation tasks, and even outperforms them on two metrics (Alignment and Overlap). Furthermore, we also extend our approach to solve two challenging problems that have more flexible constraints, namely Text-to-layout and Content-aware layout generation. The qualitative and quantitative results of our approach are superior to those of existing methods, demonstrating the effectiveness and generalization ability of LayoutPrompter on more difficult tasks, even without model training. Our code and prompts will be released.