An AI-generated technical figure can look convincing and still be a poor deliverable.
The failure usually appears after the first review: change a label, reroute an arrow, update a table, translate the annotations, or move a panel. If all that remains is a bitmap, the output is visually useful but structurally disposable.
Scientific Illustrator is an MIT-licensed open-source Codex plugin project built around a more useful target: recreating a reference figure with editable objects in PowerPoint, WPS Presentation, or draw.io. Its README says it prefers editable text, shapes, connectors, tables, and charts, keeping image inserts for the smallest areas it cannot reliably recreate, such as complex textures or microscopy-like imagery. That is a project claim, not a benchmark.
The repository structure fits that goal. It includes a plugin manifest, MCP configuration, and skills for designing, recreating, auditing, and correcting figures, plus bridge scripts for presentation software and draw.io. The plugin manifest describes a Designer–Drawer–Reviewer–Corrector workflow. The v1.5.3 release reports backend/target locking, serialized OOXML changes, and fixes around tables, charts, arrows, connectors, and exports.
The engineering idea matters even if this particular implementation is not the right tool for every team. For a structured diagram, the useful acceptance test is not only visual similarity. Ask whether text, connectors, groups, tables, and charts can survive the next human edit in the target application.
This is most relevant to researchers, technical writers, and developers who hand off editable PPTX or draw.io files. It is a weaker fit for photorealistic images, dense textures, pixel-perfect reproduction, or one-off artwork that will never be revised. The project itself acknowledges a boundary by retaining difficult regions as images.
I have not tested or run this project. This article is based on a read-only review of its public repository, README, manifest, license, and release notes. Treat compatibility, reconstruction fidelity, and workflow reliability as items to validate on your own templates and software versions. For source details, see the README, MIT license, and CI workflow.
Top comments (0)