Editable AI-Generated Research Figures: Introducing AutoFigure-Edit from Westlake University

Problem

Researchers often generate scientific diagrams as static PNG images, which cannot be edited without re‑generating the whole figure.

AutoFigure‑Edit

AutoFigure‑Edit extends the AutoFigure framework to produce fully editable vector graphics directly from paper text.

Five‑stage pipeline

Style‑conditioned image generation : a raster image is generated from the long‑form method description, optionally conditioned on a user‑provided style reference.

Segmentation & structural indexing : Meta’s SAM3 model segments visual components (icons, modules, arrows) and builds a structural skeleton.

Asset extraction : each component is extracted with a transparent background (RGBA).

SVG template generation & refinement : a layout template in SVG is created and refined using the extracted assets.

Asset injection : assets are injected into the template, yielding a layer‑wise editable SVG where every element can be moved, recolored, or relabeled.

Key innovations

Pixel‑to‑vector conversion enables direct in‑browser editing of diagrams.

Reference‑image‑guided style transfer learns color palette, fonts, and icon style from a single example.

Built‑in interactive editor allows immediate layout, text, and icon adjustments on the generated SVG.

“Reasoned Rendering” decouples logical layout from aesthetic rendering, supporting a critique‑and‑refine loop and OCR‑based replacement of blurry raster text with clean vector text.

SAM3‑driven automatic segmentation produces transparent assets for each visual component.

Experimental evaluation

Evaluation on the FigureBench benchmark shows that AutoFigure‑Edit outperforms prior methods (GPT‑Image, SVG‑Code, Diagram Agent) across accuracy, completeness, and adaptability.

Win‑rate with style guidance: 83 % vs 76 % without.

Human evaluation (217 participants, 262 generated figures):

Semantic correctness 4.04/5

Information completeness 4.11/5

Visual quality 3.95/5

Style consistency 4.09/5

48 % of figures deemed ready for paper submission without modification.

Automatic metrics (FigureBench):

Accuracy 8.83, completeness 8.26, adaptability 8.37 (significantly higher than baselines).

Without reference style: overall score 8.29, visual 8.32, expressiveness 8.66.

SVG conversion correctness 3.60/5; 36 % of users gave a perfect score, and low scores (<2) occurred in fewer than 12 % of cases.

Open‑source resources

Code and models are released under an open‑source license at https://github.com/ResearAI/AutoFigure-Edit. The FigureBench dataset is available on HuggingFace.

Impact

By delivering editable SVG diagrams, AutoFigure‑Edit removes the “final‑step” bottleneck in scientific publishing, enables consistent visual style across papers, and supports fully autonomous AI‑driven research workflows.