mPresenter

mPresenter is an end-to-end multilingual agentic framework designed to transform static academic papers (PDFs) into presentation videos.

✨ Features

• 🤖 Multi-Agent Collaboration: Orchestrates 4 specialized agents (Planner, Reviewer, Coder, Interpreter) to plan, critique, and generate professional slides.
• 👁️ Cross-Lingual Interpretation: A dedicated Interpreter Agent analyzes figures to explain visual elements that remain in the source language.
• 🎨 True-to-Layout Generation: Writes executable Beamer LaTeX code and visually inspects rendered slides to ensure high readability.
• 📈 High Information Density: Designed for effective knowledge transfer, achieving higher accuracy in QA benchmarks compared to prior systems.
• ⚡ Cost-Effective Efficiency: Achieves the lowest token consumption among baselines and substantially reduces latency, making high-quality video generation affordable and scalable.

🎬 Demo Videos

English:

mpresenter_en.mp4

中文:

mpresenter_zh.mp4

📊 mPreBench Dataset

An expert-curated multilingual benchmark designed to evaluate Effective Information Transfer (EIT) of Paper2Video systems.

• 📄 40 Academic Papers:
  • 20 English Papers (10 from NeurIPS 2025, 10 from ACL 2025).
  • 20 Chinese Papers (from Chinese Journal of Computers).
  • Covers diverse topics: Vision, NLP, Graph, Security, RL, BioMed, and Systems.
• ❓ 1,600 Multilingual Questions:
  • Each paper is paired with 8 expert-written multiple-choice questions.
  • All questions are translated into 5 languages: English, Chinese, German, Japanese, and Arabic.

Evaluation Dimensions

The benchmark targets four core aspects of scientific communication:

• Motivation: Research context and gaps in related work.
• Method: Technical mechanisms and figure interpretation.
• Experiment: Experimental setup and result analysis.
• Conclusion: Key takeaways and supported claims.

🚀 Quick Start

System dependencies:

• TeX Live (XeLaTeX).
• Poppler (pdftoppm) or ImageMagick.
• Fonts: Source Han Sans and Fira Sans

1. Install dependencies with pixi:

pixi install
# Enable CosyVoice TTS
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git third_party/CosyVoice

2. Set an LLM key:
   • OpenAI: set llm.openai_api_key in config.json.
   • Gemini: set llm.gemini_api_key in config.json or export GEMINI_API_KEY / GOOGLE_API_KEY.
3. Place your PDF at input/paper.pdf or pass --source-pdf.
4. Run:

pixi shell
python main.py --source-pdf input/paper.pdf --target-language English --output-root output

🐳 Docker Deployment

docker build -t mpresenter:latest .

Run:

mkdir -p input output cache
cp /path/to/paper.pdf input/paper.pdf
docker run --rm -it \
  -v "$(pwd)/input:/app/input" \
  -v "$(pwd)/output:/app/output" \
  -v "$(pwd)/cache:/app/cache" \
  -v "$(pwd)/config.json:/app/config.json" \
  -e GEMINI_API_KEY=YOUR_KEY \
  mpresenter:latest \
  --source-pdf /app/input/paper.pdf --target-language English --output-root /app/output

⌨️ CLI

Common command-line options:

• --source-pdf: override source_pdf (default input/paper.pdf)
• --output-root: override output_root
• --cache-root: override cache_root
• --target-language: override target_language (default: English)
• --planner-note: override planner_note
• --backbone: override all LLM model names with a single model

Other configuration is defined in config.json.

Session cache IDs are derived from the input PDF filename (stem), so re-running with the same PDF name reuses the same cache directory under cache/.

📂 Outputs

Cache directory (cache/<session_id>/) key files and folders:

• steps_status.json: pipeline step status for resume.
• run.log: main run log.
• final_outline.json: finalized outline JSON.
• slides_manifest.json: slide-level manifest (image path + note).
• slides.tex and slides.pdf: merged Beamer source and PDF.
• slides_llm/: slide PNGs for LLM review.
• slides/: slide PNGs for video synthesis.
• final_scripts.json: finalized narration scripts.
• audio/: synthesized audio clips per slide.

Final output:

• output/<source_pdf_stem>.mp4: final video.

⏱️ Efficiency Analysis

mPresenter maintains low latency and minimal API costs while producing superior video quality.

🙌 Acknowledgements

Thanks to these open-source projects:

• PaddleOCR (PP-DocLayout-L + OCR): https://github.com/PaddlePaddle/PaddleOCR
• Metropolis Beamer theme (mtheme): https://github.com/matze/mtheme
• FunAudioLLM/CosyVoice (TTS): https://github.com/FunAudioLLM/CosyVoice