Y-Research-SBU/PosterGen — 252 Stars

Topics

#agent #large-language-model #vision-language-models

Language breakdown

Python 89.2%

TypeScript 7.5%

CSS 3.0%

Shell 0.1%

HTML 0.1%

▸ Files click to expand

README

PosterGen: Aesthetic-Aware Multi-Modal Paper-to-Poster Generation
via Multi-Agent LLMs

Zhilin Zhang^{1,2 ★} Xiang Zhang^{3 ★} Jiaqi Wei⁴ Yiwei Xu⁵ Chenyu You¹

¹ Stony Brook University ² New York University ³ University of British Columbia
⁴ Zhejiang University ⁵ University of California, Los Angeles ^★ Equal Contribution

Abstract

In this work, we propose a multi-agent LLMs framework that is guided by design principles.

Our multi-agent LLMs adopt a workflow of specialist agents that mirrors a professional design process:

>

- Parser Agent – extracts and structures all content from the source paper.

- Curator Agent – designs a narrative-based storyboard.

- Layout Agent – transforms the storyboard into a spatially balanced, three-column layout.

- Styling Agents – apply a harmonious color palette and a hierarchical typographic system to ensure aesthetic coherence.

>

This methodology is designed to generate a well-designed poster that minimizes the need for manual fine-tuning.

📢 News

2025.08.26 Our paper is now available on arXiv! 📄
2025.08.23 Code Released. PosterGen now available! 🎉🎉

🚀 Quick Start

System Requirements

Operating System: Windows, Linux, or macOS
Python Version: 3.11

1. Environment Setup

Option 1: UV (Recommended)

git clone -b main https://github.com/Y-Research-SBU/PosterGen.git cd PosterGen

uv sync

Option 2: Conda

# Create and activate conda environment conda create -n poster python=3.11 -y conda activate poster git clone -b main https://github.com/Y-Research-SBU/PosterGen.git cd PosterGen

pip install -r requirements.txt

2. Install LibreOffice

Windows:

Download and install LibreOffice from official website
Add LibreOffice to your system PATH:

- Default installation: Add C:\Program Files\LibreOffice\program to PATH - Or custom installation: Add <yourinstallpath>\LibreOffice\program to PATH

macOS:

brew install --cask libreoffice

Ubuntu/Linux:

sudo apt install libreoffice Or using snap:
 sudo snap install libreoffice

3. API Keys Configuration

Create a .env file in the project root with your API keys:

OPENAIAPIKEY="youropenaikey"
ANTHROPICAPIKEY="youranthropickey"
ZHIPUAPIKEY="yourzhipukey"

To use a compatible API service or a custom deployment, set the base URL in your .env file:

OPENAIBASEURL="https://your-api-endpoint.com/v1"
ANTHROPICBASEURL="https://your-anthropic-endpoint.com"
GOOGLEBASEURL="https://your-google-endpoint.com"
ZHIPUBASEURL="https://your-zhipu-endpoint.com"

Data Structure Setup

Before running the multi-agent pipeline, organize your files in the data/ folder:

data/
└── <yourpapername>/
    ├── paper.pdf          # Your research paper (required)
    ├── aff.png           # Affiliation logo for color extraction (required)
    └── logo.png          # Conference logo for poster (required)

Examples (check `data/ folder): <pre><code class="lang-">data/ └── NeuralEncodingandDecodingat_Scale/ ├── paper.pdf ├── aff.png └── logo.png └── ...</code></pre>


🎯 Usage
Command-line Interface
Generate your poster with a single command:
For UV setup: <pre><code class="lang-bash">uv run python -m src.workflow.pipeline \   --posterwidth 54 --posterheight 36 \   --paperpath ./data/YourPaper_Name/paper.pdf \   --text_model gpt-4.1-2025-04-14 \   --vision_model gpt-4.1-2025-04-14 \   --logo ./data/YourPaperName/logo.png \   --afflogo ./data/YourPaper_Name/aff.png</code></pre>
For Conda setup: <pre><code class="lang-bash">python -m src.workflow.pipeline \   --posterwidth 54 --posterheight 36 \   --paperpath ./data/YourPaper_Name/paper.pdf \   --text_model gpt-4.1-2025-04-14 \   --vision_model gpt-4.1-2025-04-14 \   --logo ./data/YourPaperName/logo.png \   --afflogo ./data/YourPaper_Name/aff.png</code></pre>
Parameters:

--poster_width/height: Poster dimensions in inches, with aspect ratio (w/h): lower bound 1.4 (ISO A paper size), upper bound 2 (human vision limit)

--paper_path: Path to your PDF paper

--text_model: LLM for text processing (options: "gpt-4.1-2025-04-14" (default), "gpt-4o-2024-08-06", "gpt-4.1-mini-2025-04-14", "claude-sonnet-4-20250514")

--visionmodel: Vision model for analysis (same options as textmodel)

--logo: Conference logo

--aff_logo: Affiliation logo (used for color scheme extraction)

 Web Interface

> Developed by: React + TypeScript + Vite
Upload your PDF paper and logos through drag-and-drop, configure models and dimensions, then generate and download your poster files.
Prerequisites:

Node.js installed
 Main PosterGen dependencies installed (from project root)

API keys configured in .env file

 <pre><code class="lang-bash"># Install main project dependencies (if not done already, refer to Environment Setup above)
Start backend
cd webui &amp;&amp; python start_backend.py
Start frontend (in new terminal, from project root)
cd webui &amp;&amp; sh ./start_frontend.sh
Open http://localhost:3000 in your browser</code></pre>


Output Structure

After successful generation, you'll find your results in the output/ folder:


<pre><code class="lang-">output/ └── &lt;paper_name&gt;/     ├── &lt;paper_name&gt;.png           # final poster image     ├── &lt;paper_name&gt;.pptx          # editable PowerPoint file     ├── timingcostlog.json       # runtime &amp; api cost log     ├── assets/                    # extracted content from paper via Marker     │   ├── figures.json           # figure metadata with aspect ratios     │   ├── tables.json            # table metadata with aspect ratios     │   ├── figure-*.png           # individual figures from paper     │   ├── table-*.png            # individual tables from paper     │   └── figtabcaption_mapping.json  # caption mappings     └── content/                   # multi-agent artifacts         ├── raw.md                         # raw text extraction         ├── structured_sections.json      # organized sections         ├── classified_visuals.json       # categorized visuals         ├── narrative_content.json        # paper summary         ├── story_board.json              # content organization         ├── initiallayoutdata.json      # initial layout         ├── column_analysis.json          # column usage stats         ├── optimizedstoryboard.json    # balanced content         ├── balancer_decisions.json       # optimization details         ├── finalcolumnanalysis.json    # final usage metrics         ├── optimized_layout.json         # balanced layout         ├── finaldesignlayout.json      # element coordinates         ├── color_scheme.json             # color palette         ├── sectiontitledesign.json     # title styling         ├── keywords.json                  # highlighted terms         ├── styled_layout.json            # formatted text         └── styling_interfaces.json       # typography settings</code></pre>
🤖 Multi-Agent Pipeline
Our system uses 6 specialized AI agents working together:
Parser Agent: Extracts and structures content from paper PDF
Curator Agent: Plans content organization and visual placement  
Layout Agent: Calculates precise positioning and spacing
    - Balancer Sub-Agent: Optimizes column utilization and prevents overflow
Color Agent: Generates cohesive color schemes from your affiliation logo
Font Agent: Applies typography and keyword highlighting
Renderer: Generates final PowerPoint and image files

Key Features
Professional Layout: CSS-like precision positioning with proper spacing
Intelligent Balancing: Automatic column optimization prevents overflow
Color Harmony: Automatic color scheme generation from your institution branding
Typography Excellence: Professional font choices and keyword highlighting
Flexible Output: Both PNG images and editable PowerPoint files
Academic Standards: Follows poster design best practices for conferences

Other Configurations

The system supports customization through config/poster_config.yaml. You can adjust:


Layout parameters (margins, padding, spacing)
Typography settings (fonts, sizes, line spacing)
Color generation algorithms
Visual asset sizing constraints
Content optimization thresholds

Custom Fonts:
If you would like to use other fonts, you can add the font files under

fonts/, modify the getfontfilepath() mapping in src/layout/textheightmeasurement.py, and adjust the 'typography' in config/posterconfig.yaml`.

📊 Example Results

Our system generates professional academic posters with high visual quality. Here are some examples of generated posters:

Example 1

Example 2

Citation

@inproceedings{zhang2026postergen,
  title={PosterGen: Aesthetic-Aware Multi-Modal Paper-to-Poster Generation Via Multi-Agent LLMs},
  author={Zhang, Zhilin and Zhang, Xiang and Wei, Jiaqi and Xu, Yiwei and You, Chenyu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9813--9823},
  year={2026}
}

Acknowledgments

This codebase is built upon following open-source projects. We express our sincere gratitude to:

LangGraph: Multi-agent workflow framework;
Marker: High-quality PDF parsing library that enables accurate content extraction from research papers.
python-pptx: Python library for creating PowerPoint (.PPTX) poster files.