What exactly is AutoFigureEdit?

AutoFigureEdit (also referred to as auto figure edit) is a specialized tool that converts methodology descriptions from academic papers into fully customizable SVG scientific diagrams. It chains together LLM-based drafting, SAM3 component detection, and automatic vector conversion to produce camera-ready figures you can adjust at will. The underlying auto figure edit research appeared at ICLR 2026.

How does the auto figure edit pipeline operate?

The AutoFigureEdit pipeline has four stages: (1) Initial Raster Drafting -- a large language model generates a preliminary sketch from your methodology text; (2) SAM3 Component Detection -- Segment Anything Model 3 identifies and isolates distinct visual elements such as icons, labels, and arrows; (3) SVG Wireframe Construction -- structural vector frameworks are assembled with tagged placeholders; (4) Composite Assembly -- polished icons and vectorized text fill in the wireframe, producing a fully customizable SVG. This staged approach in auto figure edit guarantees consistent, high-caliber output.

How is AutoFigureEdit different from tools like DALL-E or Midjourney?

The critical distinction lies in editability. Mainstream image generators output flat raster files -- if you need to relocate one arrow or correct a label, the entire image must be regenerated. AutoFigureEdit delivers fully customizable SVG files where each element (labels, shapes, arrows, icons) can be independently selected, repositioned, scaled, and recolored. On top of that, the auto figure edit system supports reference-based style matching for visual consistency across your manuscript.

What does style matching mean in AutoFigureEdit?

The auto figure edit system can replicate the artistic direction of any reference image you provide. Upload a figure you like, and AutoFigureEdit will produce new illustrations that echo its color schemes, font choices, and visual motifs. This auto figure edit capability ensures every diagram in your publication maintains a cohesive appearance.

What role does SAM3 play in AutoFigureEdit?

SAM3 (Segment Anything Model 3) is a cutting-edge segmentation model from Meta. The auto figure edit pipeline uses SAM3 to identify and segment individual icon regions from the preliminary raster draft via multi-prompt detection. This allows dependable extraction of separate components (icons, label blocks, connectors) that are then positioned into the SVG framework as independently modifiable elements in the AutoFigureEdit output.

FigureBench is the inaugural large-scale benchmark designed for evaluating scientific illustration generation from extended text, hosted publicly on Hugging Face. It includes 3,300 samples covering papers (2,900 development + 200 test), blogs, surveys, and textbooks, with input lengths spanning 352 to 12,732 tokens. The auto figure edit system was rigorously evaluated on FigureBench across challenges like extended-context reasoning, dense text handling, and structural complexity.

What inputs does AutoFigureEdit require?

At a minimum, the auto figure edit pipeline needs the methodology section from your research paper. Beyond that, you may optionally provide a reference image for style matching, tailored SAM3 detection prompts for particular component categories, and settings for your preferred LLM provider (Gemini, OpenRouter, or other OpenAI-compatible APIs). AutoFigureEdit takes care of everything else automatically.

Can I modify the figures after generation?

Absolutely -- this is the central advantage of the auto figure edit approach. Every output is fully customizable SVG. You can use the integrated svg-edit canvas in AutoFigureEdit to refine diagrams right in your browser, or download the SVG and open it in any vector tool such as Adobe Illustrator, Inkscape, or Figma. Each component can be repositioned, resized, recolored, and altered independently. The auto figure edit workflow makes post-generation refinement seamless.

Is AutoFigureEdit open source?

AutoFigureEdit is built on the AutoFigure-Edit open-source research project hosted on GitHub under the MIT License. The open-source repository includes the core auto figure edit pipeline, SAM3 integration, and web interface. The FigureBench dataset is available on Hugging Face under CC BY 4.0. AutoFigureEdit (autofigureedit.com) is a commercial hosted platform offering a convenient cloud-based version with extra features and dedicated support.

How do AutoFigureEdit and AutoFigure-Edit relate to each other?

AutoFigure-Edit is the open-source research framework published on GitHub under the MIT License, containing the core auto figure edit pipeline (LLM drafting + SAM3 component detection + SVG vector conversion), web interface, and FigureBench benchmark. AutoFigureEdit (autofigureedit.com) is a commercial web platform built on top of AutoFigure-Edit, providing cloud-based rendering, additional capabilities, and technical support so researchers can use the auto figure edit technology without any local setup.

Which LLM providers work with AutoFigureEdit?

The auto figure edit system supports several LLM providers: Google Gemini (the default -- Gemini 3.1 Flash for image drafting, Pro for SVG generation), OpenRouter (with access to Gemini, Claude, and additional models), and any OpenAI-compatible API. You can set your preferred provider and model within the AutoFigureEdit configuration.

Accepted at ICLR 2026

AutoFigureEdit: Research Text to Modifiable Vector Figures

Turn methodology sections from academic papers into fully modifiable SVG diagrams — camera-ready scientific figures you can customize element by element.
Powered by LLM drafting, SAM3 component detection, and automatic vector conversion. Published at ICLR 2026.

Try the Generator

Read Paper (ICLR 2026)

AutoFigureEdit pipeline converting methodology text to modifiable SVG diagrams

Give AutoFigureEdit a Try

See the auto figure edit pipeline in action. Drop in your methodology text, optionally attach a reference image for style matching, and watch AutoFigureEdit build a camera-ready modifiable SVG diagram for you. The tool handles methodology visualizations, system overviews, process flowcharts, and more -- all derived from plain text and delivered as fully customizable vector artwork.

Image Generator

Method Text

tip: concise, structured method text yields cleaner templates

0 / 15000

Reference Image (optional)

Upload a sketch or diagram to enhance

tip: upload a figure from a paper you like to transfer its visual style

Cost 5 credits

Image Preview

No Images Generated

FigureBench Results

Validated on FigureBench — the first large-scale scientific illustration generation benchmark with 3,300 samples from papers, blogs, surveys, and textbooks.

3,300

Benchmark Samples

1.7k

GitHub Stars

ICLR 2026

Published At

MIT

Open Source License

AutoFigureEdit Showcase

Browse a curated collection of what the auto figure edit system can produce -- spanning methodology visualizations, system overviews, and beyond. Each sample was created through the four-stage AutoFigureEdit pipeline straight from paper text, yielding fully customizable SVG results.

Research Paper Figures

AutoFigureEdit builds camera-ready methodology visualizations straight from your research manuscript. Whether it involves intricate model designs, sequential processing stages, or encoder-decoder setups, the auto figure edit pipeline renders them as fully customizable SVG diagrams complete with precise geometry, labeled connections, and clear annotations -- ideal for top-tier venue submissions.

Survey Paper Overviews

When you need to visualize an entire research landscape, AutoFigureEdit produces detailed taxonomy and comparison charts that map relationships between multiple approaches. These auto figure edit outputs help readers rapidly understand a field's structure through organized visual hierarchies and uniform styling.

Technical Blog Graphics

Beyond formal publications, the auto figure edit pipeline works equally well with informal technical writing. Blog posts covering machine learning ideas, coding tutorials, or engineering explanations are turned into clean, visually engaging diagrams that simplify complex topics for wider audiences.

Educational Material Illustrations

AutoFigureEdit creates textbook-grade illustrations that communicate fundamental concepts with clarity. From neural network layouts and data pipeline charts to biological process flows, these auto figure edit outputs are well-suited for lecture decks, course handouts, and textbook chapters.

The Challenge AutoFigureEdit Addresses

Producing publication-grade scientific diagrams has long been a major time sink for researchers -- the auto figure edit pipeline automates the entire workflow while preserving full editability.

Tedious Manual Figure Creation

Academics routinely invest hours or entire days assembling figures in PowerPoint, Illustrator, or LaTeX tikz -- valuable time that the auto figure edit pipeline can reclaim through automation.

Locked-Down AI-Generated Images

Current AI image tools produce flat raster files with no editing capability -- a single tweak forces you to start over from scratch. AutoFigureEdit resolves this by delivering modifiable SVG output through the auto figure edit process.

Mismatched Visual Styles

Keeping a unified visual identity across every figure in a publication is notoriously difficult without style matching capabilities. AutoFigureEdit tackles this through reference-guided style transfer in the auto figure edit pipeline.

Breakthrough Contributions of AutoFigureEdit

AutoFigure-Edit brings several pioneering advances to automated scientific illustration creation, presented at ICLR 2026.

In contrast to rasterized alternatives, the auto figure edit system outputs structured Vector Graphics where every component -- labels, shapes, arrows, icons -- is independently customizable. Adjust any element losslessly with zero quality loss. No more starting from scratch for a minor revision.

Lossless vector graphics editing vs rasterized alternatives

Under the Hood: The 5-Stage Auto Figure Edit Pipeline

AutoFigureEdit converts scientific text into modifiable SVG diagrams via a 5-stage pipeline. Every stage feeds into the next, progressively transforming raw text into a fully customizable, camera-ready vector illustration.

Stage 1: Raster Drafting

A vision-language model (Gemini 3.1 Flash) processes your methodology text alongside an optional reference image, then creates an initial raster draft (figure.png). The model understands conventions of scientific figures and translates textual methodology into a coherent visual layout.

Stage 2: SAM3 Component Detection

Segment Anything Model 3 (SAM3) identifies and segments distinct visual components -- icons, label areas, connectors, shapes -- guided by structured prompts such as 'icon, person, robot, animal'. The output includes bounding boxes with confidence scores and a detection map (samed.png).

Stage 3: SVG Framework Generation

Taking the original figure, detection mask, and bounding box metadata as multimodal inputs, the LLM (Gemini 3.1 Pro) assembles a placeholder-based SVG whose regions correspond to labeled components. RMBG-2.0 strips backgrounds from cropped icons to yield transparent assets.

Stage 4: Composite Assembly

The system harmonizes coordinate systems between the SVG framework and the source figure, then swaps placeholders with transparent icons obtained from the detection stage. The result is the assembled SVG (final.svg) with every component as an independently modifiable vector element.

Stage 5: Iterative Polishing

An optional refinement stage carries out iterative SVG improvements -- path optimization, stroke recognition, and layout adjustments. Users can additionally polish results using the auto figure edit integrated svg-edit canvas with drag-and-drop composition, completing the text-to-modifiable-SVG workflow.

Questions & Answers

Everything you need to know about how AutoFigureEdit works.

Create Camera-Ready Figures in Minutes

Stop spending hours on manual diagrams. Generate publication-quality SVG figures from your methodology text with full editability — no design skills required. Start free with instant results.

Generate Your First Figure Free

Read the ICLR 2026 Paper