Figures as Programs: Recursive Generation of Editable Scientific Figures
FigTree uses recursive SVG program generation to improve scientific figure editability and quality.
Key Findings
Methodology
FigTree redefines scientific figure generation as recursive SVG program construction. Using a multi-agent system, it decomposes figures into hierarchical regions, generates local SVG code, and merges them, while employing a render-critic loop to fix visual defects.
Key Results
- FigTree achieved an overall score of 59.59 on the PaperBananaBench dataset, outperforming the strongest baseline by 9.59 points, with notable improvements in faithfulness and conciseness.
- Fine-grained diagnostics showed FigTree leading in semantic coverage and visual polish, scoring 8.39 overall.
- Experiments demonstrated that FigTree's recursive generation and critic-guided repair mechanisms significantly reduced hallucination issues.
Significance
This study addresses the pain points of poor editability and difficulty in fixing visual defects in scientific figures, providing academia with an efficient method for generating high-quality figures and advancing automated research tools.
Technical Contribution
FigTree introduces an innovative framework combining recursive decomposition and critic-guided repair, leveraging multi-agent systems and SVG code generation to enhance figure editability and quality, while offering new verification and repair mechanisms.
Novelty
FigTree is the first to redefine scientific figure generation as recursive SVG program construction, offering superior editability and semantic control compared to traditional raster-based methods.
Limitations
- Currently supports only SVG format, limiting flexibility for other vector formats.
- Efficiency for generating complex figures needs improvement.
- Render-critic loops may increase computational costs.
Future Work
Future directions include extending support for multiple vector formats, optimizing generation efficiency, and exploring smarter critic mechanisms to reduce computational overhead.
AI Executive Summary
Scientific figures are essential for academic communication, yet their creation is often labor-intensive and difficult to edit. Existing raster-based methods, while visually appealing, lack editability and are challenging to fix when defects arise.
FigTree introduces a novel recursive SVG program generation framework, decomposing figures into hierarchical regions, generating local SVG code, and merging them. A render-critic loop further refines defects, ensuring high-quality and editable outputs.
Experiments on benchmark datasets show FigTree outperforming existing methods in faithfulness, conciseness, and overall quality. This research paves the way for efficient scientific figure generation and contributes to the development of automated research tools.
Deep Analysis
Background
Scientific figures play a critical role in research papers, helping to convey complex methods clearly. Traditional figure creation relies heavily on manual design, which is time-consuming and difficult to edit. Recent advancements in raster-based image generation models like Nano Banana Pro and GPT-Image-2 have improved visual quality but fail to address editability and defect correction.
Core Problem
Existing methods produce raster figures that are hard to edit and fix when visual defects occur. Additionally, generating high-quality complex figures in a single step remains challenging, often requiring multiple iterations.
Innovation
FigTree introduces recursive decomposition to divide figures into manageable regions, generating short SVG code fragments and merging them. It also employs a render-critic loop to identify and fix defects, offering superior editability and semantic control compared to raster-based methods.
Methodology
- �� Figure decomposition: Break figures into hierarchical regions, generating independent SVG code for each.
- �� Multi-agent system: Each node is handled by a generator agent, producing local code and reporting layout information.
- �� Critic-guided repair: Render figures, inspect visual defects, trace issues to code, and refine iteratively.
Experiments
Experiments used the PaperBananaBench dataset with 292 methodology figure cases. Baselines included Nano Banana Pro and Crafter. Evaluation metrics covered faithfulness, conciseness, readability, and aesthetics.
Results
FigTree scored 46.23 in faithfulness and 64.38 in conciseness, achieving an overall score of 59.59, significantly outperforming the strongest baseline Crafter (50.00). Fine-grained diagnostics highlighted its strengths in visual polish and semantic control.
Applications
FigTree is ideal for generating academic figures, especially for complex methodologies requiring high editability. Its recursive framework can be extended to other vector graphic tasks.
Limitations & Outlook
Currently limited to SVG format, with lower efficiency for complex figures. Render-critic loops may increase computational costs. Future work could address these issues and expand format support.
Plain Language Accessible to non-experts
Imagine you're designing a complex puzzle. Traditional methods draw the entire puzzle at once, but if one piece is wrong, you have to redo the whole thing. FigTree's approach is like designing each puzzle piece individually and assembling them. If one piece is flawed, you only fix that piece without disturbing the rest.
ELI14 Explained like you're 14
Hey, imagine you're playing Minecraft! You want to build an awesome castle, but every time you mess up, you have to rebuild the whole thing. FigTree is like letting you build one room at a time and then connecting them. If one room is wrong, you just fix that room — cool, right? Makes building way easier and faster!
Glossary
SVG (Scalable Vector Graphics)
An XML-based vector graphic format offering high precision and editability.
Used as the core format for generating scientific figures.
Recursive decomposition
Breaking complex tasks into smaller, manageable subtasks.
Applied to generate hierarchical figure regions.
Render-critic loop
Inspecting rendered figures to identify and fix visual defects iteratively.
Used for refining figure quality.
Multi-agent system
A framework where multiple agents collaborate to complete complex tasks.
Used for recursive figure generation and repair.
Content hallucination
Errors where generated outputs include irrelevant or incorrect information.
FigTree reduces this issue through decomposition and critic mechanisms.
Open Questions Unanswered questions from this research
- 1 How can support for more vector formats be added?
- 2 How can generation efficiency for complex figures be improved?
- 3 How can critic loops reduce computational costs?
Applications
Immediate Applications
Academic figure generation
Rapidly produce high-quality, editable scientific figures for papers and reports.
Educational tools
Help students and educators create clear teaching diagrams efficiently.
Long-term Vision
Automated research tools
Enable end-to-end automation of research workflows, from figure generation to paper writing.
Abstract
Scientific methodology figures are essential for communicating complex methods clearly, yet creating them remains labor-intensive and typically requires multiple rounds of refinement. Recent image-generation models can synthesize visually appealing raster figures, but producing a human-satisfactory result in a single generation step remains difficult. Moreover, precise edits to raster figures are challenging for both humans and models. We formulate scientific figure generation as recursive SVG program construction and propose \textsc{FigTree}, a \textit{multi-agent} system that automatically transforms a scientific paper into a structured vector figure. \textsc{FigTree} grounds figure content in the source paper, decomposes a figure into a hierarchy of local regions, generates each region as a short SVG program, and assembles the resulting fragments. A render-critic refinement loop jointly inspects the rendered figure and its underlying program, enabling visual defects to be traced to specific statements and accurately repaired. We conduct extensive evaluations of \textsc{FigTree} on figure quality and editability, showing that \textsc{FigTree} produces high-quality figures, while also enabling more effective editing than existing raster-based methods.