Paper2Figure - A Multi-Agent Collaborative System for Figure Generation - Towards Academic Research Paper

Oct 26, 2026

Upload by Blitz

Automatically generating clear and accurate figures for research papers remains challenging because it demands semantic understanding, precise structure, and visual aesthetics. Paper2Figure introduces a dual multi-agent system with an interactive web platform for paper-to-figure generation. Generation Agents convert text into FigScript, while Refinement Agents iteratively correct logic, alignment, aesthetics, and text accuracy. Users can further refine outputs via the interface. Paper2Figure Bench evaluates 100 figures, showing +12% accuracy, +13.5% beauty, and +17% completeness without human adjustment.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Abstract

Introduction

Paper2Figure

This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. Paper2Figure: A Multi-Agent Collaborative System for Figure Generation Towards Academic Research Paper Siwei Han 1∗ Haonian Ji 1∗ Siyang Xin 1∗ Juanquan Shi 1 Shi Qiu 1 Xinyu Ye 1 Peng Xia 1 Jiaqi Liu 1 Zhaorun Chen2 Yiyang Zhou 1 Linjie Li3 Lijuan Wang3 Huaxiu Yao 1† 1UNC-Chapel Hill 2University of Chicago 3Microsoft [siweih@cs.unc.edu](siweih@cs.unc.edu) , [huaxiu@cs.unc.edu](huaxiu@cs.unc.edu) ∗Equal contribution. †Corresponding author. Abstract Automatically generating clear and accurate figures for research papers remains challenging, as it requires semantic understanding, precise structure, and visual aesthetics. Existing approaches struggle to balance fidelity and quality: large language model (LLM) code-based methods (e.g., SVG, Mermaid) are structured but inflexible, while imagegeneration models (e.g., GPT-Image-1, Nano Banana) produce hard-to-edit and often inaccurate figures. We present Paper2Figure, a dual multi-agent system with an interactive web platform for paper-to-figure generation. Generation Agents convert text into our designed FigScript language, encoding figure semantics, styles and layout. The web system renders the FigScript into an initial image, which Refinement Agents iteratively analyze to locate issues and revise the FigScript for improved logic, alignment, aestheticsand text accuracy. Crucially, users can further refine results through an intuitive web interface, ensuring full control over the final output. To evaluate Paper2Figure, we introduce Paper2Figure Bench, a benchmark comprising 100 academic figures with paired descriptions. Experiments demonstrate that Paper2Figure markedly improves accuracy by 12%, beauty by 13.5%, and completeness by 17.0% over state-of-the-art baselines infully automatic generation without human adjustment. By combining automated generation with interactive edit, Paper2Figure bridges the gap between AI assistance and researcher control, offering a practical solution for high-quality academic figure creation. 1. Introduction Effective visual communication is central to scientific research. Concepts are far easier to comprehend through structured diagrams than textual descriptions [7, 10] . Consequently, main figures in academic papers, such as process diagrams, conceptual flowcharts, and system overviews, serve as indispensable tools for explanation, and dissemination. However, creating such figures remains a laborintensive and expertise-demanding task: researchers must translate nuanced textual logic into visually coherent, aesthetically balanced, and publication-ready graphics. Recent advances in generative AI have inspired two main directions for automatic figure generation, each with fundamental limitations. The first relies on large language models (LLMs) to produce structured code, such as Scalable Vector Graphics (SVG) [5, 6, 24, 25, 37] or markup languages like Mermaid [37] . These approaches provide semantic clarity and editability but often yield messy layouts and poor aesthetics, producing figures that are structurally correct yet visually unappealing. The second direction leverages powerful text-to-image models [16, 26, 30–32, 36, 41], including GPT-Image-1 [23] and Nano Banana (Gemini 2.5 Flash Image) [9], to synthesize richly styled figures. Despite their visual appeal, these models struggle with layout accuracy, textual rendering, and editability, making their outputs impractical for scientific use [4, 14, 17, 34] . As a result, no existing method can jointly ensure semantic precision, visual quality, and flexible structure control. We propose Paper2Figure, a dual multi-agent system [11, 38] with an interactive web platform for automatic paper-to-figure generation. Paper2Figure introducesa structur

FAQs

What problem does Paper2Figure address in academic paper figure creation?

It addresses the difficulty of generating figures that are semantically correct, structurally sound, and visually aesthetic while remaining editable and publication-ready.

How does Paper2Figure generate and refine figures?

Generation Agents translate text into FigScript, which the web system renders into an initial image. Refinement Agents analyze the image, identify issues, and revise FigScript to improve logic, alignment, aesthetics, and text accuracy.

How is Paper2Figure evaluated and what improvements are reported?

The paper introduces Paper2Figure Bench with 100 academic figures and paired descriptions. Results show improvements over state-of-the-art baselines: +12% accuracy, +13.5% beauty, and +17% completeness under fully automatic generation without human adjustment.

Oct 26, 2026

Upload by Blitz