Repolode
Submit
potamides

potamides/DeTikZify

A multimodal LLM that turns scientific figures and hand sketches into compilable TikZ/LaTeX graphics programs.

View on GitHubWebsite
Stars
1,828
Forks
93
Open issues
17
Last commit
2 months ago

What is it?

DeTikZify is a multimodal language model that synthesizes TikZ graphics programs from sketches and existing figures. The goal is to make it easier to produce scientific figures from a rough drawing, or to recreate a figure that exists only as an image and lacks its underlying semantic source.

The project provides a Python package with a pipeline API for sampling a single TikZ program, rasterizing compiled output, and running an MCTS-based inference procedure that iteratively refines results over a set time budget without additional training. It also supports text-conditioned generation through the TikZero adapters and TikZero+ models, and includes a web UI launchable from the command line.

Using the package requires a full TeX Live 2023 installation, ghostscript, and poppler. Model weights and the DaTikZ datasets are hosted on the Hugging Face Hub, and a hosted demo is available in a Hugging Face Space.

Key features

  • 01Generates TikZ programs from sketches or figure images.
  • 02Compiles and rasterizes generated TikZ code for preview.
  • 03Uses MCTS-based inference to refine outputs without extra training.
  • 04Supports text-conditioned figure generation via TikZero adapters.
  • 05Includes a command-line launchable web UI and runnable example scripts.

Use cases

  • Turning a hand-drawn diagram into editable LaTeX figure code for a paper.
  • Recreating a scientific figure that exists only as an image.
  • Experimenting with text-to-TikZ generation for academic illustrations.

Quick start

  1. 1

    Install the Python package with pip

    pip install 'detikzify[legacy] @ git+https://github.com/potamides/DeTikZify'
  2. 2

    Install a full TeX Live 2023, ghostscript and poppler through your package manager

  3. 3

    Start the web UI from the command line

    python -m detikzify.webui --light

Things to know

  • Requires a full TeX Live 2023 installation plus ghostscript and poppler on top of the pip package.
  • The README notes the free Colab tier only supports the 1b models, so the 8b models need substantial hardware.
  • Text-conditioned generation via TikZero is only available through the programming interface, not the web UI.

Topics

#tikz#latex#multimodal#llm#scientific-figures#sketch-to-code#vision-language-model#draw#graph#huggingface

FAQ

What is DeTikZify?
A multimodal LLM that turns scientific figures and hand sketches into compilable TikZ/LaTeX graphics programs. DeTikZify is a multimodal language model that synthesizes TikZ graphics programs from sketches and existing figures. The goal is to make it easier to produce scientific figures from a rough drawing, or to recreate a figure that exists only as an image and lacks its underlying semantic source.
Is DeTikZify free to use? What license does it have?
DeTikZify is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is DeTikZify still maintained?
Updated within 6 months. The last commit to DeTikZify was on Sep 17, 2026.
What can I use DeTikZify for?
Turning a hand-drawn diagram into editable LaTeX figure code for a paper. Recreating a scientific figure that exists only as an image. Experimenting with text-to-TikZ generation for academic illustrations.
How do I install DeTikZify?
Install the Python package with pip: pip install 'detikzify[legacy] @ git+https://github.com/potamides/DeTikZify' Install a full TeX Live 2023, ghostscript and poppler through your package manager Start the web UI from the command line: python -m detikzify.webui --light
Where can I get DeTikZify?
The source code and installation instructions are on GitHub: https://github.com/potamides/DeTikZify

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects