논문 그림에 AI 챗봇 활용하기 (2025.1)

English summary

This presentation describes practical workflows and prompting strategies for using large language models and specialized AI tools to plan and draft scientific figures and figure legends. It frames AI as a tool for early brainstorming, structured planning, visual prototyping, and legend drafting rather than as an autonomous producer of final publication graphics.

Key points

• Standard text to image diffusion models generate images by matching descriptive text with visual pixel patterns, whereas scientific figures require precise geometric placement, coordinate based arrangement of shapes and text, explicit connections such as arrows, and highly abstract symbolic representation.
• Direct image generation can produce anatomical inaccuracies, nonsensical chemical structures, unreadable text and labels, and incorrect number sequences. The presentation cites retracted papers containing AI generated graphics with biologically impossible rat anatomy and human knee diagrams with anatomical errors.
• OpenAI ChatGPT o1 can use manuscripts or review drafts to propose figure titles, visual layouts and flows, shape arrangements, links to key findings, and suggested colors and labels. When draft figures and manuscript text are supplied, it can produce structured, publication style legends describing experimental conditions, statistical representations, and panel specific content. The presentation also describes asking the model to design an effective legend prompt in advance. The model is described as having greater latency and computational cost because of its internal reasoning process and as supporting text and multiple image inputs.

Audience

Researchers, biomedical scientists, academic authors, and students preparing scientific manuscripts, figures, figure legends, and conceptual diagrams.

Scope and limitations

The source is based on practical presentation evidence notes and does not report a systematic comparative evaluation or independent validation of publication quality. Claude 3.5 Sonnet with Artifacts is described as separating chat from a rendering workspace and supporting iterative work with SVG, HTML, Markdown, Mermaid, and React components. It can draft vector figures in SVG, generate Mermaid flowcharts, and revise layouts and visual styles conversationally. Napkin AI is described as transforming unstructured text into editable diagrams and visual metaphors, representing topics such as pharmacological mechanisms in different layouts and exporting SVG or PDF files. However, current LLMs and AI chatbots are not presented as capable of autonomously producing camera ready, publication grade figures or making fine graphic adjustments. The recommended workflow is to export prototype layouts or SVG and Mermaid code, then manually finalize layout, aesthetics, and domain accuracy in dedicated software such as Adobe Illustrator or Microsoft PowerPoint.

Research context

This topic has a neutral thematic connection to Sangzin Ahn's portfolio through the use of AI in medical research, scientific visualization, and data science based research communication.