VILA-Lab/FigMirror

What it solves

FigMirror allows users to recreate the visual style of a scientific figure from a research paper and apply it to their own data. It eliminates the tedious process of manually editing plotting code to match a specific aesthetic, providing an editable matplotlib script and a camera-ready PDF as the final output.

How it works

The system uses an agentic loop consisting of a Drawer and a Reviewer. An Orchestrator delegates the task of rendering a figure to the Drawer agent, which produces a candidate figure using "Grounded Measurement" (using x/y coordinates and executable checks to ensure visual accuracy). The Reviewer agent then compares the candidate against the reference image, providing structured feedback with bounding boxes and a revision checklist. This iterative process continues until the output matches the reference style, with a "preserve list" used to prevent style drift across iterations.

Who it’s for

Researchers and scientists who need to produce high-quality, publication-ready figures that match the style of existing academic papers.

Highlights

• Reference-Conditioned Style Transfer: Generates figures based on a reference image and user-provided data.
• Agentic Loop: Employs a role-separated Drawer/Reviewer architecture to refine visual output.
• Grounded Measurement: Uses foundation models to return coordinates and execute pixel-level checks for precision.
• Multi-Backend Support: Available as skills for Claude Code and Codex, with an optional local Web UI.
• 3D Figure Support: Includes geometry-aware prompting for camera, scale, and lighting to preserve 3D compositions.