PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation

TL;DR

PosterO reframes content-aware poster layout generation as a layout-tree problem expressed in SVG, enabling large language models to perform layout prediction through in-context learning with intent-aligned examples. By vectorizing both shapes and design intents and encoding hierarchical dependencies, PosterO achieves data-efficient, generalized layout generation that adapts to diverse poster purposes. Extensive experiments on PKU PosterLayout, CGL, and the newly introduced PStylish7 demonstrate state-of-the-art performance and robustness to domain shifts, while ablations validate the importance of intent vectors, hierarchical structure, and example selection. The work also provides a new benchmark for generalized content-aware layout generation, highlighting practical potential for real-world design tasks and suggesting directions for interactive, feedback-driven improvements.

Abstract

In poster design, content-aware layout generation is crucial for automatically arranging visual-textual elements on the given image. With limited training data, existing work focused on image-centric enhancement. However, this neglects the diversity of layouts and fails to cope with shape-variant elements or diverse design intents in generalized settings. To this end, we proposed a layout-centric approach that leverages layout knowledge implicit in large language models (LLMs) to create posters for omnifarious purposes, hence the name PosterO. Specifically, it structures layouts from datasets as trees in SVG language by universal shape, design intent vectorization, and hierarchical node representation. Then, it applies LLMs during inference to predict new layout trees by in-context learning with intent-aligned example selection. After layout trees are generated, we can seamlessly realize them into poster designs by editing the chat with LLMs. Extensive experimental results have demonstrated that PosterO can generate visually appealing layouts for given images, achieving new state-of-the-art performance across various benchmarks. To further explore PosterO's abilities under the generalized settings, we built PStylish7, the first dataset with multi-purpose posters and various-shaped elements, further offering a challenging test for advanced research.

Figures (19)

• Figure 1: An overview of PosterO. First, (a) takes image-layout pairs $(I, L)$ from datasets as input for data preparation, jointly modeling various-shaped elements $E$ and design intents $D$ towards layout trees $T$ and building up the latent space of $D$. Then, (b) takes test images $I_t$ as input and searches examples based on the predicted design intents $f_t$ to apply LLM $\mathcal{M}$ through in-context learning. After obtaining generated layout trees $\hat{T}$, (c) can continue the conversation with $\mathcal{M}$ to create poster designs seamlessly.
• Figure 2: Various shapes of elements in PStylish7 dataset.
• Figure 3: Comparisons of visualized results on the unannotated test split of PKU PosterLayout dataset.
• Figure 4: Comparisons of visualized results on the unannotated test split on CGL dataset.
• Figure 5: Content metrics over in-context example size $k$.