Paper2Poster/Paper2Poster

What it solves

Creating a poster from a research paper is slow and manual. Paper2Poster automates the whole process: it takes a paper.pdf and produces an editable poster.pptx, and it also provides tools to objectively judge how good a generated poster is.

How it works

The core engine is PosterAgent, a top-down, visual-in-the-loop multi-agent system. It works in three stages:

1. Parser – Distills the paper into a structured asset library (text, figures, tables, etc.).
2. Planner – Aligns text–visual pairs into a binary-tree layout that preserves reading order and spatial balance.
3. Painter-Commentor loop – Each poster panel is refined by executing rendering code, then a vision-language model (VLM) gives feedback to fix overflow and alignment issues.

You can mix and match LLMs and VLMs (e.g. GPT-4o or local Qwen-2.5-7B models), and run generation in parallel with --max_workers. The system also supports YAML-based style customization, automatic conference/institution logo search (local store + DuckDuckGo or Google), Docker deployment, and a Gradio demo.

For evaluation, Paper2Poster includes a dataset and three metric families: PaperQuiz (question-answering about the paper based on the poster), VLM-as-Judge, and statistical measures. You can also create a PaperQuiz for your own paper.

Who it’s for

• Researchers and students who need conference posters without manual layout work.
• Developers or AI agents (like Codex or Claude) who want a lightweight skill to prepare poster-ready content from a paper.
• Anyone working on automated document-to-visual generation, especially in the academic domain.

Highlights

• Generates editable .pptx posters directly from a paper PDF.
• Multi-agent pipeline with visual feedback for high-quality layout.
• Flexible model combination: use cloud APIs or local open-source models (via vLLM).
• Automatic logo insertion for conferences and institutions.
• YAML-based style customization and a default theme.
• Evaluation suite: PaperQuiz, VLM-as-Judge, and statistical metrics.
• Docker and Gradio support for easy deployment and demos.
• A lightweight agent SKILL for use with coding assistants.