AI Scientific Illustration Tips from a PhD Researcher

The quality of a research paper depends heavily on its illustrations. A clear methodology diagram helps reviewers instantly grasp your approach, while a poorly drawn figure can hurt your paper's acceptance chances.

Yet most graduate students receive no formal training in scientific visualization. Learning Adobe Illustrator or TikZ takes significant time — time better spent on actual research.

During an academic sharing session, a PhD researcher from the Center for Fire Safety Materials at Beijing University of Chemical Technology shared practical experience using AI tools for scientific illustration. Here are the key insights and actionable tips from that talk.

Choosing the Right AI Tool for Research Figures

Not all AI image generators are suitable for academic illustrations. The researcher compared the most popular options:

The researcher's verdict:

NanoBanana is an AI illustration tool designed for academic scenarios. Compared to general-purpose tools, it offers noticeable advantages in generation quality and logical consistency for scientific figures. She also noted that NanoBanana is a paid tool, but there are free alternatives available such as ZenMux, LOVART, and LMArena.

PaperBanana: Taking It Further Than NanoBanana

PaperBanana builds significantly on NanoBanana's core philosophy, providing researchers with a more powerful toolset:

5-Agent Multi-Agent Architecture

PaperBanana employs a unique architecture with 5 specialized AI agents working together:

• Retriever Agent: Searches a database of figures from top venues (NeurIPS, ICLR, etc.) to find similar references, ensuring outputs follow academic conventions
• Planner Agent: Breaks down your methodology text into structured visual elements and layout plans
• Stylist Agent: Optimizes colors, fonts, and typography based on the target journal/conference style
• Visualizer Agent: Transforms the plan into high-quality image output
• Critic Agent: Scores each generated result and suggests improvements, driving iterative refinement

Iterative Optimization

NanoBanana typically generates results in a single pass. PaperBanana's Critic Agent automatically runs multiple "generate → review → improve" cycles until publication-quality is achieved. You simply input your text, and the system handles multi-round optimization automatically.

Two Core Tools

The Prompt Engineering Formula

Writing effective prompts is the key to getting good results. The researcher shared a practical formula:

Subject Description + Environment/Background + Art Style + Render Engine + Color Scheme + Aspect Ratio

What each element does:

• Subject: What you want to draw — nanoparticles, molecular structures, experimental setups
• Environment: Context — white background, lab setting, microscopic view
• Art Style: Cutaway diagram, schematic, vector style, 3D render
• Material: Metallic, translucent, glass texture
• Lighting: Studio lighting, octane render
• Aspect Ratio: 1:1 square, 16:9 wide, 4:3 standard

Prompt Examples

Example 1: Nanoparticle schematic

Example 2: Thermal management diagram

Example 3: Methodology flowchart

Paste your paper's methodology section directly into PaperBanana's AI Diagram Generator — the AI analyzes the text and produces a structured methodology diagram automatically, with no need to write complex prompts manually.

Three Real-World Use Cases

Case 1: Refining Existing Figures

The most common need — you have a rough figure that needs polishing or modification.

The researcher demonstrated a nanomaterials example: she took an existing figure and used AI to add a glowing halo effect, making the key elements stand out.

How to do it: Upload your original image to PaperBanana's AI Image Editor and describe your desired changes in natural language. For example: "Add a golden glow halo to this nanoparticle and change the background to pure white."

Case 2: Merging Multiple Figures

When you need to combine the best aspects of two figures into one, AI can help.

She showed a thermal management example: one figure had good overall layout but lacked detail, while another had the needed detail. She found that general-purpose Chinese AI tools produced logically inconsistent merges, while academic-focused tools generated more accurate and better-composed results.

PaperBanana's 5-Agent architecture is especially advantageous for this type of task — the Critic Agent checks the logical consistency of merged outputs, preventing results that "look good but are scientifically inaccurate."

Case 3: Generating Figures from Scratch

When you have no starting point, this is where AI shines most.

Recommended workflow:

1. Paste the abstract or methodology text into the AI Diagram Generator
2. The 5 AI agents collaborate to generate an initial figure
3. The Critic Agent automatically runs multiple rounds of iterative optimization
4. Fine-tune the final result further using the AI Image Editor if needed
5. Use the final result as reference or inspiration for further refinement

This method works especially well for graphical abstracts and mechanism diagrams — two types of figures that graduate students find most challenging.

Using AI-Generated Figures in Papers

The researcher also addressed critical considerations when using AI-generated images in publications:

AI Detection and Watermarks

AI-generated images typically contain three types of identifiers:

1. Visible watermarks: Some tools add obvious brand logos or text watermarks in the corners of images
2. Invisible watermarks (Digital Steganography): Digital signatures hidden in color distributions, completely invisible to the human eye but detectable through frequency domain analysis. This is the most important type to be aware of — even if an image looks clean, AI detection tools can still identify it
3. Metadata watermarks (Metadata / C2PA): An "ID card" embedded in the image's EXIF data, recording the generation model, prompt, and creation time. These follow industry standards like C2PA with authenticity-proving file headers

Recommended Best Practices

Rather than trying to remove these identifiers, the recommended approach is to:

• Use AI output as reference and inspiration, then redraw using tools like Illustrator
• Transparently disclose AI usage in your paper — top venues like NeurIPS and ICML increasingly require authors to declare AI tool usage
• Use specialized academic tools like PaperBanana — compared to general-purpose AI, academic-focused tools generate figures with better scientific logic and formatting conventions, making them more suitable for direct use in publications
• Post-process with vector software (e.g., Adobe Illustrator) — convert AI output to editable vector format for fine-tuning. Note that vectorization may reduce image quality; you can adjust the trace color count settings higher to preserve more original detail

Get Started with PaperBanana

Ready to try AI-assisted scientific illustration? You can start with PaperBanana's two core tools:

• AI Diagram Generator: Paste your methodology text and the 5-Agent architecture automatically generates publication-ready academic diagrams
• AI Image Editor: Upload existing figures and describe changes in natural language

See pricing plans for more details.

This article is based on a public academic sharing session by a PhD researcher from the Center for Fire Safety Materials at Beijing University of Chemical Technology. The tool comparisons and usage recommendations represent the speaker's personal experience. We encourage readers to choose tools based on their own research needs and institutional AI usage policies.