AI Scientific Figure Generator

🎨 Generate publication-quality scientific figures using AI models

This tool uses a sophisticated two-step workflow to create NeurIPS/ICLR-style scientific figures from your research text:

1. GPT-5 analyzes your scientific text and generates a structured MODULE LIST
2. Gemini-2.5-flash-image (nano banana) creates a professional figure based on the MODULE LIST

✨ Features

• 🤖 Integrated AI Workflow: GPT-5 analysis guides nano banana for optimal results
• 📊 Publication-Ready: Generates clean, NeurIPS-style scientific figures
• 🎯 Structured Approach: Two-step process ensures logical, accurate visualizations
• 🔧 Easy Configuration: Simple .env file setup
• 💾 Automatic Saving: Saves both MODULE LIST and generated figures
• 🖥️ CLI Interface: User-friendly command-line tool

🚀 Quick Start

1. Installation

# Clone the repository
git clone git@github.com:hengzzzhou/FigForge.git
cd FigForge

# Install dependencies
pip install -r requirements.txt

2. Configuration

Create a .env file from the template:

cp .env.example .env

Edit .env and add your API credentials:

Option 1: OpenAI-Compatible API (default, for relay endpoints)

API_TYPE=openai
OPENAI_BASE_URL=
OPENAI_API_KEY=
ANALYSIS_MODEL=gpt-5
IMAGE_MODEL=gemini-2.5-flash-image
OUTPUT_DIR=outputs

Option 2: Native Google Gemini API

API_TYPE=gemini
GEMINI_API_KEY=
OPENAI_BASE_URL=
OPENAI_API_KEY=
ANALYSIS_MODEL=gpt-5
IMAGE_MODEL=gemini-2.5-flash-image
OUTPUT_DIR=outputs

Note: When using API_TYPE=gemini, the MODULE LIST generation still uses the OpenAI-compatible API, but image generation uses the native Google Gemini API.

3. Usage

Generate from a file:

python scientific_plotter.py -i examples/sample_input.txt

Generate from text directly:

python scientific_plotter.py -t "Your scientific text describing your model architecture..."

Specify output location:

python scientific_plotter.py -i input.txt -o my_awesome_figure.png

Generate MODULE LIST only:

python scientific_plotter.py -i input.txt --module-list-only

📸 Examples

Here are some scientific figures generated using FigForge:

Sample Input - Neural Architecture

LiveSearchBench

ReSo

VIKI-R

All figures are generated with clean conference-style design, featuring flat aesthetics, consistent line weights, and professional color palettes.

---

📖 How It Works

Step 1: MODULE LIST Generation (GPT-5)

The GPT-5 model analyzes your scientific text and creates a structured MODULE LIST that breaks down your architecture into:

1. Input(s): Data sources and preprocessing
2. Preprocessing/Encoding/Embedding: Feature extraction layers
3. Core Architecture/Stages/Blocks: Main model components in sequence
4. Special Mechanisms: Attention, memory, routing, etc.
5. Output Head: Final prediction layers

Step 2: Figure Generation (Gemini-2.5-flash-image)

Using the MODULE LIST as a guide, nano banana generates a clean, professional figure following these design principles:

• ✅ Flat, clean conference style (no gradients, shadows)
• ✅ Consistent thin line weights
• ✅ Professional pastel color palette
• ✅ Rounded rectangles for module blocks
• ✅ Clear arrows indicating data flow
• ✅ Concise labels (no long sentences)
• ✅ Pure white background with clean spacing

📁 Project Structure

NBP/
├── scientific_plotter.py      # Main application script
├── requirements.txt            # Python dependencies
├── .env.example               # Environment variable template
├── .gitignore                 # Git ignore patterns
├── prompts/
│   ├── step1_module_generation.txt    # Prompt for MODULE LIST
│   └── step2_figure_generation.txt    # Prompt for figure generation
├── examples/
│   └── sample_input.txt       # Sample scientific text
├── outputs/                   # Generated figures (auto-created)
└── README.md                  # This file

🎯 Example Workflow

$ python scientific_plotter.py -i examples/sample_input.txt

================================================================================
🚀 Starting Scientific Figure Generation Workflow
================================================================================

📊 Step 1: Generating MODULE LIST using gpt-5...
✅ MODULE LIST generated successfully!

================================================================================
MODULE LIST:
================================================================================
[Structured breakdown of your architecture...]
================================================================================

📝 MODULE LIST saved to: outputs/sample_input_module_list_20231125_143022.txt

🎨 Step 2: Generating figure using gemini-2.5-flash-image...
💾 Downloading image from: [URL]
✅ Figure saved to: outputs/sample_input_20231125_143022.png

================================================================================
🎉 Workflow completed successfully!
================================================================================
📄 MODULE LIST: outputs/sample_input_module_list_20231125_143022.txt
🖼️  Figure: outputs/sample_input_20231125_143022.png
================================================================================

⚙️ Configuration Options

Environment Variables

| Variable | Description | Default | |----------|-------------|---------| | API_TYPE | API type for image generation: openai or gemini | openai | | OPENAI_BASE_URL | OpenAI-compatible API endpoint URL | Required | | OPENAI_API_KEY | API key for OpenAI-compatible endpoint | Required | | GEMINI_API_KEY | Google Gemini API key (required when API_TYPE=gemini) | - | | ANALYSIS_MODEL | Model for MODULE LIST generation | gpt-5 | | IMAGE_MODEL | Model for figure generation | gemini-2.5-flash-image | | OUTPUT_DIR | Directory for output files | outputs |

Command-Line Options

| Option | Short | Description | |--------|-------|-------------| | --input FILE | -i | Path to input text file | | --text TEXT | -t | Scientific text as string | | --output FILE | -o | Custom output path | | --module-list-only | | Generate MODULE LIST only | | --help | | Show help message |

🔍 Tips for Best Results

1. Provide Clear Text: The more detailed and structured your input text, the better the MODULE LIST
2. Describe Flow: Explicitly mention data flow and connections between components
3. Specify Components: Name specific layers, blocks, or mechanisms in your architecture
4. Review MODULE LIST: Check the generated MODULE LIST before proceeding to figure generation
5. Iterate: You can regenerate figures with modified MODULE LIST for fine-tuning

🐛 Troubleshooting

Problem: Error initializing plotter

• Solution: Make sure .env file exists and contains valid API credentials

Problem: FileNotFoundError: Prompt template not found

• Solution: Ensure prompts/ directory exists with both template files

Problem: API connection errors

• Solution: Verify your OPENAI_BASE_URL and OPENAI_API_KEY are correct

Problem: Image generation fails

• Solution: Check that gemini-2.5-flash-image model is available on your API endpoint

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

📝 License

This project is open source and available under the MIT License.

🙏 Acknowledgments

• Prompt templates based on NeurIPS/ICLR figure design principles
• Powered by OpenAI-compatible API endpoints
• Uses GPT-5 for analysis and Gemini-2.5-flash-image (nano banana) for generation

---

Happy Scientific Plotting! 🎨✨

For questions or issues, please open an issue on GitHub.