What is naturepanelforge?

NaturePanelForge is a research tool that takes a scientific chart image from a published paper and tries to produce working Python code that reproduces that chart. The target audience is researchers who want to adapt or reuse a figure style from a Nature-family journal paper, or who want to build benchmark datasets of scientific plotting tasks. The workflow has three main stages. First, it retrieves the full figure from an open-access paper and splits compound figures (those with panels labeled A, B, C, etc.) into individual crops. Second, an AI model classifies each panel by chart type, bar chart, bubble plot, line graph, heatmap, and so on, and scores how well-suited it is for code reproduction. Third, an AI agent loop writes Python and matplotlib code to reproduce the panel, renders the result, checks it against a set of criteria, and refines the code over several attempts if the output does not match. You can use the tool in four modes. The simplest is to run a pre-written script that rerenders a checked-in example. The second installs a skill into an AI coding agent (such as OpenAI Codex) so you can paste a prompt describing the target image and have the agent handle the full reproduction process. The third mode is a full pipeline that starts from a paper DOI, fetches the figures, splits them, classifies them, and runs the agent loop in sequence. The fourth is a web gallery for browsing and searching completed reproductions. The tool is oriented toward scientific computing and reproducibility research. It is not a general chart editor or a tool for creating new charts from scratch. If you have a specific reference figure style you want to apply to your own data, the README notes that a different project called FigMirror is a better fit for that task. NaturePanelForge is focused on turning existing published panels into verifiable, rerunnable code as a benchmark exercise.