Paper2Poster is an open-source project whose goal is to automatically generate academic posters from research papers, while also addressing the question of "how to evaluate posters." At its core is PosterAgent—a top-down, visual-in-the-loop multi-agent system that can convert paper.pdf into an editable poster.pptx.
Main capabilities and features:
- Multi-agent poster automation: generate editable PPTX posters from paper PDFs.
- Supports parallel generation of section content for speedup (specified via --max_workers).
- Automatically adds logos for conferences and institutions, supports YAML style customization, and provides a new default theme.
- Supports both local deployment (via vLLM) and API-based access (such as GPT-4o), allowing flexible combinations of LLM / VLM.
- Provides a Gradio demo and Docker support.
- Provides a lightweight Paper2Poster SKILL for AI agents such as Codex and Claude to prepare poster content without relying on a full environment.
- Comes with an evaluation dataset and the PaperQuiz evaluation method, which can use PaperQuiz, VLM-as-Judge, and other statistical metrics (such as visual similarity, PPL, etc.) to evaluate generated posters.
Usage: For installation, the project provides two approaches: local deployment and Docker deployment. The local environment requires a Python environment, LibreOffice, poppler, and creating a .env file in the project root directory to add the OpenAI API key; for more reliable logo search, an optional Google Search API can be configured (add GOOGLE_SEARCH_API_KEY and GOOGLE_SEARCH_ENGINE_ID to .env). Docker deployment can be run by building the image. When using it, create a folder named after the paper in the dataset directory and place a file named paper.pdf in it. If using an open-source model, it must first be deployed with vLLM, and the correct port must be specified in get_agent_config() in utils/wei_utils.py. Example configurations include: [High-performance] use GPT-4o to generate posters; [Economical] use Qwen-2.5-7B-Instruct combined with GPT-4o; [local] uses Qwen-2.5-7B-Instruct. Logo addition supports a local logo library (logo_store/institutes/ and logo_store/conferences/) with priority lookup; if not found, it performs a web search, using DuckDuckGo by default (no API key required), and Google Custom Search can be used via --use_google_search; a custom logo path can also be specified directly to skip automatic detection. Style customization is done through YAML configuration: the global default config/poster.yaml applies to all posters; each poster can place a poster.yaml next to paper.pdf to override it. For evaluation, the Paper2Poster evaluation dataset can be downloaded, with papers stored in the Paper2Poster-data directory, and PaperQuiz, VLM-as-Judge, or other statistical metrics can be used to evaluate generated posters; a PaperQuiz can also be created for your own paper.
Versions and status:
- [2025.5.27] Released arXiv, code, and dataset.
- [2025.9.3] Added support for parallel generation of section content.
- [2025.9.18] Accepted by the NeurIPS 2025 Dataset and Benchmark Track.
- [2025.10.13] Added automatic logo, YAML style customization, and a new default theme.
- [2025.10.18] Added Docker support.
- [2025.11.3] Added Gradio demo support.
- [2025.10.7] Released the follow-up work Paper2Video.
- [2026.6.6] Launched the lightweight Paper2Poster SKILL.
Applicable scenarios: Suitable for researchers to quickly convert papers into conference posters, and also suitable for researchers to use benchmarks such as PaperQuiz to evaluate poster generation quality. The project thanks CAMEL, OWL, Docling, and PPTAgent for the codebases they provided.
Limitations: Specific charges, prices, platform support scope, generation results, etc. are not clearly stated in the provided materials and are not currently available.
