Paper2Demo: Turn any academic PDF into a stunning demo video — fully automatically.

Drop in a research paper. Walk away with a production-quality video presentation.

⚠️ Early version — This is a v0 prototype. Expect rough edges. Contributions and feedback are very welcome.

---

🔧 How It Works

The overall architecture of the Paper2Demo pipeline.The system transforms a raw PDF into a dynamic video via three main modules: local spatial extraction (extractor.py), AI-driven semantic structuring and summarization (analyzer.py), and dynamic video rendering with spring-physics animations (Remotion)

---

🎥 Demo

OBS-Diff: Accurate Pruning for Diffusion Models in One-Shot

OBS-DIFF--ACCURATE-PRUNING-FOR-DIFFUSION.mp4

A real video abstract generated entirely by Paper2Demo.

---

✨ What it does

Given a PDF like this:

Attention Is All You Need (Vaswani et al., NeurIPS 2017)

It produces a ~30-second animated video with:

• Title scene — paper title, authors, venue, adaptive color theme
• Problem scene — the challenge the paper addresses, with animated bullet cards
• Method scene — pipeline diagram + the paper's own architecture figure
• Contributions scene — key contributions in a clean grid layout
• Results scene — quantitative stats + the paper's results figure

Every video is automatically themed based on the paper's domain — blue for efficiency/compression, purple for generative models, teal for NLP, orange for RL, and more.

---

🚀 Quick Start

1. Clone and install

git clone https://github.com/yourusername/paper2demo
cd paper2demo

# Remotion (video renderer)
npm install

# Python pipeline
cd paper2demo
pip install -r requirements.txt

2. Get an API key

DeepSeek is recommended — cheap, fast, and works globally. Analyzing a full paper costs roughly $0.01.

export DEEPSEEK_API_KEY=sk-...

# Or use Google Gemini
export GEMINI_API_KEY=AIza...

3. Run

cd paper2demo
python main.py /path/to/paper.pdf

4. Preview

cd ..
npm run dev
# Open http://localhost:3000 → select PaperDemo

5. Export

npx remotion render PaperDemo

That's it. One command, one video.

---

🎨 Adaptive Themes

The AI detects the paper's domain and applies a matching color theme automatically.

Domain	Theme	Palette
Model Compression / Pruning	efficiency	Blue · Purple
Image Generation / Diffusion	creative	Purple · Pink
NLP / Large Language Models	language	Teal · Green
Reinforcement Learning	action	Orange · Amber
Computer Vision	vision	Sky Blue · Indigo
Multimodal (Vision-Language)	fusion	Cyan · Magenta
Robotics	robotics	Steel · Orange
Medical / Bioinformatics	medical	Blue · Aqua
Theory / Optimization	academic	Gold · Purple
Systems / Infrastructure	systems	Emerald · Blue

---

📁 Project Structure

paper2demo/              Python pipeline
├── main.py              CLI entry point
├── extractor.py         PDF parsing — text + figures (PyMuPDF)
├── analyzer.py          AI analysis → paper-data.json
└── requirements.txt

src/PaperDemo/           Remotion video composition
├── index.tsx            Main composition + calculateMetadata
├── types.ts             TypeScript types for paper-data.json
├── shared.tsx           Reusable animation hooks + components
└── scenes/
    ├── TitleScene.tsx
    ├── ProblemScene.tsx
    ├── MethodScene.tsx          ← embeds extracted figure
    ├── ContributionsScene.tsx
    └── ResultsScene.tsx         ← embeds extracted figure

public/paper/            Auto-generated (gitignored)
├── paper-data.json
└── fig_*.png

---

⚙️ CLI Options

python main.py paper.pdf [--provider deepseek|gemini] [--output /path/to/project]

Flag	Default	Description
--provider	deepseek	AI provider (deepseek or gemini)
--output	parent of paper2demo/	Remotion project root

---

💡 Tips

Cost — A full 22-page paper costs ~$0.01 with DeepSeek. The chunked pipeline makes even 100-page papers affordable.

Figure quality — Figures are extracted at full resolution from the PDF. The AI matches each figure to the right scene by reading caption text, not by guessing filenames.

Speed — Extraction is instant (local). Analysis takes 30–60 seconds depending on paper length and provider. Rendering takes ~2 minutes on a modern laptop.

---

🗺️ TODO List

• Direct arXiv URL input (python main.py arxiv:2312.00752)
• Multiple video templates (minimal white, magazine, cinematic)
• AI-generated voiceover via ElevenLabs
• Batch processing for reading lists
• Auto-generate animated charts from paper tables
• Web UI — drag and drop, no terminal needed

---

🤝 Contributing

PRs are welcome. If you make a video from a cool paper, share it and tag the repo.

---