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

• 📄 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.

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

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

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}

--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." 

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

# 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

# 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:

💾 Checkpoint & Resume:

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

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

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

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

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:

• 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).

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) 

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

❤️ Thanks for visiting ✨ Paper2Slides!