Paper2Slides: From Paper to Presentation in One Click

✨ Never Build Slides from Scratch Again ✨

| 📄 Universal File Support  |  🎯 RAG-Powered Precision  |  🎨 Custom Styling  |  ⚡ Lightning Speed |

---

🎯 What is Paper2Slides?

Turns your research papers, reports, and documents into professional slides & posters in minutes.

✨ Key Features

• 📄 Universal Document Support

Seamlessly process PDF, Word, Excel, PowerPoint, Markdown, and multiple file formats simultaneously.

• 🎯 Comprehensive Content Extraction

RAG-powered mechanism ensures every critical insight, figure, and data point is captured with precision.

• 🔗 Source-Linked Accuracy

Maintains direct traceability between generated content and original sources, eliminating information drift.

• 🎨 Custom Styling Freedom

Choose from professional built-in themes or describe your vision in natural language for custom styling.

• ⚡ Lightning-Fast Generation

Instant preview mode enables rapid experimentation and real-time refinements.

• 💾 Seamless Session Management

Advanced checkpoint system preserves all progress—pause, resume, or switch themes instantly without loss.

• ✨ Professional-Grade Visuals

Deliver polished, presentation-ready slides and posters with publication-quality design standards.

⚡ Easy as One Command

# One command to generate slides from a paper
python -m paper2slides --input paper.pdf --output slides --style doraemon --length medium --fast --parallel 2

---

🔥 News

• [2025.12.09] Added parallel slide generation (--parallel) for faster processing
• [2025.12.08] Paper2Slides is now open source!

---

🎨 Custom Styling Showcase

doraemon

academic

custom

doraemon

academic

custom

_{✨ Multiple styles available — simply modify the --style parameter}

Examples from DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

💡 Custom Style Example: Totoro Theme

--style "Studio Ghibli anime style with warm whimsical aesthetic. Use soft watercolor Morandi tones with light cream background, muted sage green and dusty pink accents. Totoro character can appear as a friendly guide relating to the content, with nature elements like soft clouds or leaves."

---

🌐 Paper2Slides Web Interface

---

📋 Table of Contents

• 🎯 Quick Start
• 🏗️ Paper2Slides Framework
• 🔧 Configuration
• 📁 Code Structure

---

🏃 Quick Start

1. Environment Setup

# Clone repository
git clone https://github.com/HKUDS/Paper2Slides.git
cd Paper2Slides

# Create and activate conda environment
conda create -n paper2slides python=3.12 -y
conda activate paper2slides

# Install dependencies
pip install -r requirements.txt

[!NOTE] Create a .env file in paper2slides/ directory with your API keys. Refer to paper2slides/.env.example for the required variables.

2. Command Line Usage

# Basic usage - generate slides from a paper
python -m paper2slides --input paper.pdf --output slides --length medium

# Generate poster with custom style
python -m paper2slides --input paper.pdf --output poster --style "minimalist with blue theme" --density medium

# Fast mode
python -m paper2slides --input paper.pdf --output slides --fast

# Enable parallel generation (2 workers by default)
python -m paper2slides --input paper.pdf --output slides --parallel 2

# List all processed outputs
python -m paper2slides --list

CLI Options:

Option	Description	Default
--input, -i	Input file(s) or directory	Required
--output	Output type: slides or poster	poster
--content	Content type: paper or general	paper
--style	Style: academic, doraemon, or custom	doraemon
--length	Slides length: short, medium, long	short
--density	Poster density: sparse, medium, dense	medium
--fast	Fast mode: skip RAG indexing	false
--parallel	Enable parallel slide generation: --parallel uses 2 workers, --parallel N uses N workers	1 (sequential without this option)
--from-stage	Force restart from stage: rag, summary, plan, generate	Auto-detect
--debug	Enable debug logging	false

💾 Checkpoint & Resume:

Paper2Slides intelligently saves your progress at every key stage, allowing you to:

Scenario	Command
Resume after interruption	Just run the same command again — it auto-detects and continues
Change style only	Add --from-stage plan to skip re-parsing
Regenerate images	Add --from-stage generate to keep the same plan
Full restart	Add --from-stage rag to start from scratch

[!TIP] Checkpoints are auto-saved. Just run the same command to resume. Use --from-stage only to force restart from a specific stage.

3. Web Interface

Launch both backend and frontend services:

./scripts/start.sh

Or start services independently:

# Terminal 1: Start backend API
./scripts/start_backend.sh

# Terminal 2: Start frontend
./scripts/start_frontend.sh

Access the web interface at http://localhost:5173 (default)

---

🏗️ Paper2Slides Framework

Paper2Slides transforms documents through a 4-stage pipeline designed for reliability and efficiency:

Stage	Description	Checkpoint	Output
🔍 RAG	Parse documents and construct intelligent retrieval index using RAG	checkpoint_rag.json	Searchable knowledge base
📊 Analysis	Extract document structure, identify key figures, tables, and content hierarchy	checkpoint_summary.json	Structured content map
📋 Planning	Generate optimized content layout and slide/poster organization strategy	checkpoint_plan.json	Presentation blueprint
🎨 Creation	Render final high-quality slides and poster visuals	Output directory	Polished presentation materials

💾 Smart Recovery System

Each stage automatically saves progress checkpoints, enabling seamless resumption from any point if the process is interrupted—no need to start over.

Fast Mode vs Normal Mode

Mode	Processing Pipeline	Use Cases
Normal	Complete RAG indexing with deep document analysis	Complex research papers, lengthy documents, multi-section content
Fast	Skip RAG indexing, direct LLM query	Short documents, instant previews, quick revisions

Use --fast when: - Document (text + figures) is short enough to fit in LLM context - Quick preview/iteration needed - Don't want to wait for RAG indexing

Use normal mode (default) when: - Document is long or has many figures - Multiple files to process together - Need retrieval for better context selection

---

⚙️ Configuration

Output Directory Structure

outputs/
├── <project_name>/
│   ├── <content_type>/                   # paper or general
│   │   ├── <mode>/                       # fast or normal
│   │   │   ├── checkpoint_rag.json       # RAG query results & parsed file paths
│   │   │   ├── checkpoint_summary.json   # Extracted content, figures, tables
│   │   │   ├── summary.md                # Human-readable summary
│   │   │   └── <config_name>/            # e.g., slides_doraemon_medium
│   │   │       ├── state.json            # Current pipeline state
│   │   │       ├── checkpoint_plan.json  # Content plan for slides/poster
│   │   │       └── <timestamp>/          # Generated outputs
│   │   │           ├── slide_01.png
│   │   │           ├── slide_02.png
│   │   │           ├── ...
│   │   │           └── slides.pdf        # Final PDF output
│   │   └── rag_output/                   # RAG index storage
│   └── ...
└── ...

Checkpoint Files: | File | Description | Reusable When | |------|-------------|---------------| | checkpoint_rag.json | Parsed document content | Same input files | | checkpoint_summary.json | Figures, tables, structure | Same input files | | checkpoint_plan.json | Content layout plan | Same style & length/density |

Style Configuration

Style	Description
academic	Clean, professional academic presentation style
doraemon	Colorful, friendly style with illustrations
custom	Any text description for LLM-generated style

Image Generation Providers

• Set IMAGE_GEN_PROVIDER in paper2slides/.env to choose the backend:
• openrouter (default): uses IMAGE_GEN_API_KEY, IMAGE_GEN_BASE_URL, and IMAGE_GEN_MODEL (default google/gemini-3-pro-image-preview)
• google: uses the official Gemini API at GOOGLE_GENAI_BASE_URL (default https://generativelanguage.googleapis.com/v1beta), IMAGE_GEN_API_KEY, IMAGE_GEN_MODEL (default models/gemini-3-pro-image-preview, must be image-capable), and IMAGE_GEN_RESPONSE_MIME_TYPE (default text/plain; use text types if your model does not support image responses)
• Reference figures are sent as inline data when supported (Google) or as image_url attachments (OpenRouter).

Image Generation Notes

[!TIP] By default Paper2Slides uses gemini-3-pro-image-preview (OpenRouter) for image generation; you can switch to an image-capable Google Gemini model (e.g., models/gemini-1.5-flash) via IMAGE_GEN_PROVIDER=google. Key findings:

• Mood Keywords: Words like "warm", "elegant", "vibrant" strongly influence the overall color palette
• Layout vs Style: Fine-grained layout instructions ground well; fine-grained element styling does not
• Prompt Length: Simple prompts generally outperform detailed ones
• Multi-slide Generation: Native multi-image output is story-like; for consistent slides, we use iterative single-image generation

---

📁 Code Structure

Module	Description
paper2slides/core/	Pipeline orchestration, 4-stage execution
paper2slides/raganything/	Document parsing & RAG indexing
paper2slides/summary/	Content extraction: figures, tables, paper structure
paper2slides/generator/	Content planning & image generation
api/	FastAPI backend for web interface
frontend/	React frontend (Vite + TailwindCSS)

Click to expand full project structure

Paper2Slides/
├── paper2slides/                 # Core library
│   ├── main.py                   # CLI entry point
│   ├── core/
│   │   ├── pipeline.py           # Main pipeline orchestration
│   │   ├── state.py              # Checkpoint state management
│   │   └── stages/
│   │       ├── rag_stage.py      # Stage 1: Parse & index
│   │       ├── summary_stage.py  # Stage 2: Extract content
│   │       ├── plan_stage.py     # Stage 3: Plan layout
│   │       └── generate_stage.py # Stage 4: Generate images
│   │
│   ├── raganything/
│   │   ├── raganything.py        # RAG processor
│   │   └── parser.py             # Document parser
│   │
│   ├── summary/
│   │   ├── paper.py              # Paper structure extraction
│   │   └── extractors/           # Figure/table extractors
│   │
│   ├── generator/
│   │   ├── content_planner.py    # Slide/poster planning
│   │   └── image_generator.py    # Image generation
│   │
│   ├── prompts/                  # LLM prompt templates
│   └── utils/                    # Utilities
│
├── api/server.py                 # FastAPI backend
├── frontend/src/                 # React frontend
└── scripts/                      # Shell scripts (start/stop)

---

🙏 Related Open-Sourced Projects

• LightRAG: Graph-Empowered RAG
• RAG-Anything: Multi-Modal RAG
• VideoRAG: RAG with Extremely-Long Videos

---

**🌟Found Paper