Creating scientific figures is often a time-consuming part of academic research. Researchers need to convert complex methodologies, architectures and experimental processes into diagrams that are both understandable and visually consistent. A new AI research project called AutoFigure aims to reduce that workload by automatically generating scientific illustrations from natural-language descriptions and long-form scientific text, including research papers. The project was presented as an ICLR 2026 paper and introduces an agentic approach that plans, generates, evaluates and refines scientific figures before producing the final illustration.
What Is AutoFigure?
AutoFigure is an AI-powered system designed to generate scientific illustrations from text. Instead of manually drawing a methodology diagram, system architecture or scientific workflow, users can provide a description or a research paper and allow the system to identify the important concepts and relationships before generating a figure.
The research team describes AutoFigure as an agentic framework for automatically generating high-quality scientific illustrations from long-form scientific text. The system performs multiple stages of reasoning, recombination and validation rather than simply converting individual sentences into images.
The project is available as open-source software through GitHub, and the researchers have also released the FigureBench dataset for evaluating scientific illustration generation.
From Research Paper to Scientific Figure
One of AutoFigure’s most useful capabilities is its paper-to-figure workflow.
The project’s documentation shows that users can provide a PDF or Markdown research paper and ask AutoFigure to generate a scientific figure based on the paper’s methodology. The system extracts relevant methodological information and uses it to construct the illustration.
This can be particularly useful for papers containing complex:
- Machine-learning architectures
- Research methodologies
- Scientific workflows
- System architectures
- Multi-stage pipelines
- Encoder-decoder structures
- Process diagrams
Rather than starting with a blank canvas, researchers can use the generated figure as a starting point and refine it according to their requirements.
AutoFigure Uses an Iterative Generation and Review Process
A key difference between AutoFigure and a simple text-to-image generator is its generation and evaluation loop.
According to the project’s documentation, AutoFigure first generates an initial SVG or XML representation. An evaluation component then reviews the result and provides feedback. The system can subsequently refine the figure based on that feedback. This generate-evaluate-refine process is designed to improve both the structure and visual quality of the resulting illustration.
The basic workflow can therefore be described as:
Research text → Figure generation → Evaluation → Refinement → Final scientific illustration
This approach is particularly important for scientific diagrams because simply producing an attractive image is not enough. The figure also needs to represent the relationships and methodology contained in the source material.
FigureBench: A Benchmark for Scientific Illustration Generation
The researchers behind AutoFigure also introduced FigureBench, a benchmark designed specifically to evaluate AI-generated scientific illustrations from long-form text.
The dataset contains 3,300 text-figure pairs, including approximately 3,200 research papers, along with examples from blogs, surveys and textbooks. The project describes FigureBench as a large-scale benchmark for evaluating scientific illustration generation from long-form scientific text.
This benchmark is important because evaluating scientific image generation is more complicated than judging whether an image simply looks attractive. A useful scientific figure must capture the important information from the source text while maintaining an understandable structure.
AutoFigure Supports Editable Outputs
Another important aspect of the project is its use of structured vector formats.
The original AutoFigure project supports outputs including SVG and mxGraph XML, with the latter being compatible with tools such as draw.io. This makes the generated diagrams more useful than a conventional raster image because the underlying figure can be further edited.
The project’s later AutoFigure-Edit work takes this concept further by converting scientific text into structured, editable SVG illustrations. The system is designed so that individual components of the generated figure can be modified rather than requiring the entire image to be regenerated.
AutoFigure-Edit Takes the Project Further
The AutoFigure project has evolved beyond the original text-to-figure system.
AutoFigure-Edit focuses specifically on generating editable scientific illustrations from research text. It combines language-model-based figure generation with component detection, SVG generation and an embedded editor. The research was presented as an ACL 2026 system demonstration.
The system is intended to address an important weakness of many AI-generated images: even when the visual result looks good, editing individual elements can be difficult.
With an SVG-based workflow, researchers can modify text, shapes, arrows and other components more easily.
Reference Images Can Help Control the Style
AutoFigure-Edit also supports reference-guided styling. Users can provide a reference image that helps guide the visual appearance of the generated scientific illustration. The system is designed to balance three objectives: semantic faithfulness to the source text, stylistic consistency with the reference image and structural decomposition that supports later editing.
This could be useful for researchers who want figures across a paper or research group to follow a consistent visual style.
How AutoFigure Works
At a high level, the original AutoFigure system follows an agentic workflow rather than a single image-generation step.
The system first interprets the scientific text and determines what should appear in the illustration. It then generates the initial figure and evaluates the result. Feedback from the evaluation stage can be used to refine the layout and improve the final output.
The newer AutoFigure-Edit pipeline adds additional stages for converting generated imagery into structured editable vector graphics. The project describes a workflow involving image generation, component segmentation, SVG templating, final assembly and optional iterative refinement.
Who Can Use AutoFigure?
AutoFigure can be particularly useful for people who regularly need to communicate complex technical information visually.
Researchers and Scientists
Researchers can use the system to create initial methodology diagrams, system architectures and scientific workflows for papers and presentations.
Students
Students working on dissertations, theses and research projects can use AI-generated figures as a starting point for explaining complex methodologies.
Academic Authors
Authors preparing research papers can use the system to transform lengthy methodology descriptions into visual representations.
Technical Teams
The same approach can also be useful for technical documentation where complex processes need to be communicated through diagrams.
However, researchers should carefully verify AI-generated figures against the original research. A generated illustration should be treated as an assisted visualization rather than an automatic guarantee of scientific correctness.
Why AI-Generated Scientific Figures Matter
Scientific communication depends heavily on visual representation. A well-designed diagram can make a complicated methodology considerably easier to understand.
The problem is that creating such diagrams manually can take significant time, particularly when a paper contains multiple interconnected components.
AutoFigure’s approach attempts to automate part of this process by allowing AI to interpret long-form scientific text and construct a visual representation. Its evaluation and refinement stages are intended to improve the structural and aesthetic quality of the generated figures. The researchers report that AutoFigure consistently outperformed the baseline systems evaluated on FigureBench.
Limitations Researchers Should Consider
Despite its potential, AutoFigure should not be treated as a replacement for human review.
Scientific figures contain information that can be highly sensitive to small errors. An incorrect label, missing relationship or misunderstood methodology can change the meaning of a figure.
Researchers should therefore compare the generated diagram with the original paper or methodology before using it in a publication.
There is also an important distinction between “AI-generated” and “publication-ready.” AutoFigure’s research reports strong results and describes its output as publication-ready, but individual figures may still require editing and verification depending on the complexity of the source material and the requirements of a particular journal or conference.
Conclusion
AutoFigure is an interesting example of how AI is moving beyond text generation into scientific communication and visualization. The ICLR 2026 project uses an agentic workflow to generate and refine scientific illustrations from text and research papers, while its FigureBench benchmark provides a way to evaluate this emerging capability.
The project’s evolution into AutoFigure-Edit also addresses an important practical requirement: generated scientific figures should not only look good but should be editable. By producing structured SVG output, the system gives researchers greater control over the final diagram.
For researchers, the biggest opportunity is not necessarily eliminating figure design completely. Instead, tools such as AutoFigure could provide a strong first draft, allowing scientists to spend less time drawing diagrams and more time reviewing, correcting and communicating their research.
Official project: AutoFigure on GitHub
Research paper: AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations
Disclaimer: AI-generated scientific figures should be independently reviewed for factual, methodological and visual accuracy before being used in research papers, presentations or other academic publications.