Research map / Large language models / Agents, code and math
PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PosterO, a tuning-free approach that represents poster layouts as SVG layout trees with design-intent vectorization and prompts LLMs (Llama 2, CodeLlama, Llama 3.1) with intent-aligned in-context examples, and introduces the PStylish7 dataset for generalized layout generation.
Problem
Content-aware poster layout generation suffers from limited training data, monotonous rectangular element representations, and poor handling of shape-variant elements and diverse design intents.
Approach
Represent layouts as hierarchical SVG trees that encode arbitrary-shaped elements and design-intent regions (detected by a U-Net-based intent model), then use LLM in-context learning with examples selected by intent-embedding similarity, and realize posters via follow-up chat with the LLM.
Key results
- On PKU PosterLayout unannotated split, PosterO (Llama 3.1-8B) reaches Avg 0.0677 vs 0.1434 for RALF, a 52.8% reduction in the average metric.
- On CGL annotated test split, PosterO (Llama 3.1-8B) reaches Avg 0.0203 vs 0.0238 for RALF and 0.0602 for PosterLlama.
- On PKU, PosterO outperforms LayoutPrompter with the same backbone across Llama 2-7B, CodeLlama-7B and Llama 3.1-8B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PosterO instead of PosterLlama | Content-aware layout generation · PKU PosterLayout annotated test | Avg: 0.0891 -> 0.0470 | Helped |
| Use PosterO instead of PosterLlama | Content-aware layout generation · PKU PosterLayout unannotated test | Avg: 0.0983 -> 0.0677 | Helped |
| Use PosterO instead of PosterLlama | Content-aware layout generation · CGL annotated test | Avg: 0.0602 -> 0.0203 | Helped |
| Use PosterO instead of PosterLlama | Content-aware layout generation · CGL unannotated test | Avg: 0.4359 -> 0.2612 | Helped |
| Use 7 or 8 intent-aligned in-context examples rather than 10 (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.0677 -> 0.0663; 0.0677 -> 0.0651 | Helped |
| Add hierarchical node representation (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.1078 -> 0.0677 | Helped |
| Add design intent vectorization (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.2260 -> 0.0677 | Helped |
| Add design intent vectorization and hierarchical node representation (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.2447 -> 0.0677 | Helped |
| Select examples by intent embeddings rather than intent bounding boxes (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.0836 -> 0.0677 | Helped |
| Select examples by intent embeddings rather than element types (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.1304 -> 0.0677 | Helped |
| Select examples by intent embeddings rather than randomly (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.1293 -> 0.0677 | Helped |
| Use Llama 3.2-1B rather than Llama 3.1-8B (ablation) | PosterO · PKU PosterLayout unannotated test | Avg: 0.0677 -> 0.1467 | Hurt |
Surprising findings
- Removing design intent vectorization or hierarchical nodes can improve content metrics, but this is attributed to invalid empty underlays inflating utilization.
- Random or element-type-based example selection still outperforms most baselines, suggesting even weak example alignment helps.
Limitations
- Performance on the PStylish7 generalized dataset is acknowledged as far from saturated, with substantial room for improvement.
- Relies on an intent detection model trained on existing e-commerce poster datasets, whose transfer to other purposes is not fully established.
Experimental setup
Datasets: PKU PosterLayout CGL PStylish7
Benchmarks: PKU PosterLayout CGL PStylish7
Metrics: Ove Ali Undl Unds Int Sal Rea Avg Cov Con
Baselines: CGL-GAN DS-GAN ICVT LayoutDM RALF PosterLlama LayoutPrompter
Builds on: LayoutPrompter PosterLlama RALF SegFormer
Closest papers by meaning
- Rethinking The Training And Evaluation of Rich-Context Layout-to-Image Generation (NeurIPS 2024 · similarity 0.87)
- Seeing is Improving: Visual Feedback for Iterative Text Layout Refinement (CVPR 2026 · similarity 0.87)
- POSTA: A Go-to Framework for Customized Artistic Poster Generation (CVPR 2025 · similarity 0.87)
- AesthetiQ: Enhancing Graphic Layout Design via Aesthetic-Aware Preference Alignment of Multi-modal Large Language Models (CVPR 2025 · similarity 0.87)
- Visual Layout Composer: Image-Vector Dual Diffusion Model for Design Layout Generation (CVPR 2024 · similarity 0.86)
- Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.