AutoFigure is an LLM-driven system designed to automate the creation and refinement of publication-ready scientific illustrations ICLR 2026. It addresses the challenge of generating accurate and aesthetically pleasing diagrams from text descriptions or research papers, targeting researchers and academics seeking to streamline their publication workflow. The primary benefit is the automated generation of high-quality figures, reducing manual effort and improving visual communication.

How It Works

The core of AutoFigure operates on a "Review-Refine loop" employing a dual-agent architecture. An initial generation agent produces an SVG or mxGraph XML figure based on input text or paper content. A separate critic agent then evaluates the generated figure's quality, providing specific feedback. This process iterates, with the generation agent refining the figure based on the critic's input, until a predefined quality threshold or maximum iteration count is met, ensuring publication-grade output.

Quick Start & Requirements

Installation is recommended via the Python SDK: clone the repository, navigate to the directory, and run pip install -e .. A prerequisite is playwright install chromium for rendering. Users will need API keys for supported LLM providers (e.g., OpenRouter, Gemini, Bianxie) for generation and enhancement. An interactive Next.js web interface is also available via ./start.sh, accessible at http://localhost:6002.

Highlighted Details

• Text-to-Figure: Generates diagrams directly from natural language descriptions.
• Paper-to-Figure: Extracts methodologies from PDF or Markdown documents to create visual diagrams.
• Iterative Refinement: Employs a dual-agent (Generation + Evaluation) loop for continuous quality optimization.
• Output Formats: Supports SVG and mxGraph XML (draw.io compatible).
• AI Image Enhancement: Offers optional AI-powered post-processing for aesthetic beautification with customizable art styles.
• FigureBench Dataset: Introduces a large-scale benchmark for evaluating scientific illustration generation models (HuggingFace).

Maintenance & Community

Community support is available via a WeChat discussion group. For inquiries or expired QR codes, users can contact tuchuan@mail.hfut.edu.cn.

Licensing & Compatibility

The project is released under the MIT License, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The system requires API keys for LLM providers, which may incur usage costs. The effectiveness of generation and refinement is dependent on the quality of the input description or paper content and the capabilities of the underlying LLM. No specific limitations regarding platform support or known bugs are detailed in the README.