AutoFigure is a dual-agent system that generates publication-ready scientific illustrations from text descriptions or research papers. AutoFigure-Edit (the latest version) outputs fully editable SVGs with SAM3 segmentation and style transfer.

Is there a free hosted version?

Yes. deepscientist.cc is free for all scholars. No installation required — paste your method text, optionally provide a style reference, and get an editable SVG.

How does style transfer work?

Provide a reference figure from another paper, and AutoFigure-Edit will mimic that visual style while generating your diagram. Useful for maintaining consistency across a paper or matching a journal's aesthetic.

What output formats does AutoFigure support?

AutoFigure-Edit outputs SVG (fully editable vector graphics) and PNG. The SVG output means every component — text, shapes, arrows — can be modified losslessly in any vector editor.

How do I install AutoFigure-Edit?

Docker is recommended: copy .env.example to .env, set your HF_TOKEN, run 'docker compose up -d --build', then open localhost:8000. CLI and source installation also available.

Can AutoFigure generate figures from research papers?

Yes. Paper-to-Figure mode extracts methodology from a PDF and creates visual diagrams automatically. Useful for literature reviews and visual summaries.

What is SAM3 and why does it matter?

SAM3 (Segment Anything 3) detects and segments icons, diagrams, and text regions in generated figures. This enables converting raster drafts into structured SVGs where each element is individually editable.

AutoFigure: Generate Publication-Ready Scientific Diagrams from Text

Q: What is AutoFigure-Edit?

AutoFigure-Edit is the next version of AutoFigure. It outputs editable SVGs instead of just raster images, uses SAM3 for icon detection, supports style transfer to match reference papers, and includes an embedded SVG editor.

By Prahlad Menon Published 2026-03-17 2 min read

Update (March 2026): AutoFigure-Edit is now available with editable SVG output, SAM3 segmentation, style transfer, and a free hosted version at deepscientist.cc. The paper was accepted to ICLR 2026.

Creating figures for research papers is tedious. You know what you want to show, but translating that into a polished diagram means hours in draw.io, PowerPoint, or—if you’re ambitious—TikZ.

AutoFigure generates publication-ready scientific illustrations from text descriptions or directly from research papers.

What’s New in AutoFigure-Edit?

The latest version introduces major upgrades:

Feature	Original AutoFigure	AutoFigure-Edit
Output format	SVG + mxGraph XML	Editable SVG (every element selectable)
Icon handling	Basic	SAM3 segmentation (precise region detection)
Style control	None	Style transfer (match any reference figure)
Editor	External (draw.io)	Embedded SVG editor in browser
Hosted version	None	deepscientist.cc (free)

How the New Pipeline Works

AutoFigure-Edit transforms method text into editable SVGs in four stages:

Method Text → Gemini generates draft PNG
           → SAM3 segments icons/regions  
           → Converts to SVG template
           → Injects vectorized icons
           → Editable final.svg

Stage 1: Generation — The LLM renders a journal-style schematic from your method description.

Stage 2: Segmentation — SAM3 detects distinct icons, text regions, and diagram elements. Overlapping detections are merged intelligently.

Stage 3: Templating — The system constructs an SVG wireframe with labeled placeholders matching each detected region.

Stage 4: Assembly — High-quality cropped icons (background-removed via RMBG-2.0) are injected into the template, producing a fully editable vector file.

Style Transfer: Match Any Paper’s Aesthetic

This is the killer feature. Provide a reference figure from a paper you like, and AutoFigure-Edit mimics that visual style:

Same color palette
Similar icon styles
Matching layout patterns
Consistent typography feel

Useful for:

Maintaining visual consistency across your paper
Matching a journal’s expected figure style
Replicating a style you admire from another paper

Try It Free: deepscientist.cc

No installation required. The hosted version at deepscientist.cc is free for all scholars:

Paste your method section text
(Optional) Upload a style reference image
Generate → Edit → Download SVG

The embedded editor lets you tweak text, reposition elements, and adjust colors before exporting.

Self-Hosted Installation

Docker (Recommended)

git clone https://github.com/ResearAI/AutoFigure-Edit.git
cd AutoFigure-Edit

# Configure
cp .env.example .env
# Edit .env: set HF_TOKEN=hf_xxx (for RMBG-2.0 access)

# Build and run
docker compose up -d --build

# Open http://localhost:8000

CLI Usage

pip install -r requirements.txt

# Install SAM3 separately
git clone https://github.com/facebookresearch/sam3.git
cd sam3 && pip install -e .

# Generate
python autofigure2.py \
  --method_file paper.txt \
  --output_dir outputs/demo \
  --provider bianxie \
  --api_key YOUR_KEY

Original Features (Still Available)

Two Generation Modes

Text-to-Figure — Describe what you want:

"A flowchart showing the transformer architecture with 
attention mechanism, feed-forward layers, and residual connections"

Paper-to-Figure — Upload a PDF, and AutoFigure extracts the methodology and creates visual diagrams automatically.

AutoFigure still uses its dual-agent approach:

Generate — LLM creates initial figure
Evaluate — Critic agent scores quality (0-10)
Refine — Loop until publication-ready

This Review-Refine loop is why outputs are actually usable, unlike most AI diagram generators.

When to Use AutoFigure vs Alternatives

Tool	Best For	Output
AutoFigure-Edit	Method diagrams, pipeline figures	Editable SVG
draw.io	Manual diagram creation	SVG/PNG
Mermaid	Code-defined flowcharts	SVG
TikZ	LaTeX-native, precise control	PDF
ChatGPT/Claude	Quick sketches (no editing)	PNG only

AutoFigure-Edit hits the sweet spot: AI-generated starting points that you can actually edit and refine.

Frequently Asked Questions

What is AutoFigure-Edit?

AutoFigure-Edit is the next version of AutoFigure, accepted to ICLR 2026. It generates fully editable SVG figures from method text using SAM3 segmentation and Gemini, with style transfer and an embedded browser-based editor.

Is the hosted version really free?

Yes. deepscientist.cc is free for all scholars with no usage limits announced.

What APIs/models does it use?

Gemini (via OpenRouter or direct) for image generation and SVG creation. SAM3 (Segment Anything 3) for icon detection. RMBG-2.0 for background removal.

Can I use my own LLM?

The system supports multiple providers including OpenRouter, Bianxie, and direct Gemini API. Check the configuration docs for switching providers.

How accurate is the style transfer?

It captures overall aesthetic — colors, icon styles, layout patterns — rather than pixel-perfect replication. Results vary by reference complexity. Works best with clean, well-structured reference figures.

Links: