ERNIE-Image

ERNIE-Image for Scientific Visualization & Academic Illustration: Complete Guide from Concept to Publication-Ready Figures

Jul 5, 2026

ERNIE-Image for Scientific Visualization & Academic Illustration: Complete Guide from Concept to Publication-Ready Figures

Abstract: Research papers, conference posters, textbook illustrations — high-quality academic figures are essential for every researcher. With a LongTextBench score of 0.9733 and powerful complex layout generation, ERNIE-Image is the top open-source model for academic illustration. This guide covers everything from prompt design to ComfyUI batch workflows, giving you a complete scientific visualization pipeline.

If you're a researcher, graduate student, or science educator, you've likely experienced this frustration: you spend hours creating a perfect experimental workflow diagram, only to have your advisor say "the labels are too small to read." Or you generate a molecular structure figure, and reviewers point out "the arrow direction is wrong." You submit a paper, and it's sent back for revision because the figures don't meet publication standards.

Traditional solutions are either expensive (professional illustration outsourcing costs thousands of dollars) or time-consuming (hand-drawing in Illustrator or PowerPoint takes hours). AI image generation is changing this landscape — and ERNIE-Image happens to be one of the best open-source models for scientific visualization.

Why ERNIE-Image? Three core reasons:

  1. Precise text rendering: LongTextBench score of 0.9733 (with PE), accurately rendering labels in both Chinese and English
  2. Strong complex layout capability: Officially demonstrated 4×6 grid charts, multi-panel compositions, and infographics
  3. Open-source and free for private deployment: Apache 2.0 license, academic institutions can deploy locally with data staying on campus

I. Main Types of Academic Illustrations

Scientific visualization covers many types, each with different model capability requirements:

1. Molecular Structures and Chemistry Diagrams

Molecular structure diagrams require precise lines, correct bond angles, and clear atom labels. ERNIE-Image's text rendering capability allows it to directly annotate chemical formulas on generated molecular diagrams.

Prompt Example:

A professional molecular structure diagram of aspirin (acetylsalicylic acid),
showing the benzene ring, carboxylic acid group, and acetyl group.
Clean line art, white background, labeled atoms in black, academic illustration style,
publication-ready quality, 2D structural formula

2. Cell Biology and Pathway Diagrams

Signaling pathway diagrams, cell structure figures, protein interaction networks — these are common illustration types in research papers. ERNIE-Image excels at handling complex multi-element layouts.

Prompt Example:

A detailed cell signaling pathway diagram showing the MAPK cascade,
with labeled proteins (Ras, Raf, MEK, ERK), phosphorylation arrows,
nucleus and cytoplasm compartments, clean academic style,
color-coded components, white background, publication quality

3. Experimental Workflow and Comparison Diagrams

"Two-row comparison" figures are among the most common illustration types in academic papers: showing traditional method vs. new method, before vs. after treatment, control group vs. experimental group.

Prompt Example:

A two-row comparison scientific diagram. Top row labeled "Conventional Workflow":
manual process with hand-drawn icons. Bottom row labeled "AI-Enhanced Workflow":
automated process with digital icons. Clean arrows connecting steps,
numbered stages (1-6), academic illustration style, white background

4. Data Visualization and Infographics

AI-generated versions of bar charts, radar charts, heatmaps. While ERNIE-Image cannot replace professional data visualization software (R/ggplot2), it can generate conceptual visualizations for paper abstracts, posters, and science communication articles.

Prompt Example:

A clean scientific bar chart comparing five AI image generation models,
showing GenEval scores from 0.80 to 0.90, labeled axes,
color-coded bars, title "Model Performance Comparison",
white background, academic publication style, clear labels in English

5. Physics and Engineering Schematic Diagrams

Mechanical analysis diagrams, circuit designs, optical systems — ERNIE-Image performs well on structured engineering illustrations.

Prompt Example:

A technical engineering diagram showing a photovoltaic solar cell cross-section,
labeled layers (anti-reflection coating, n-type silicon, p-type silicon, back contact),
arrows showing electron flow, clean schematic style, white background,
publication-ready technical illustration

II. Core Prompt Design Techniques

Technique 1: Specify Style Keywords

Academic illustrations have a fixed style system. Adding these keywords to your prompt enhances professionalism:

  • academic illustration style — Academic style
  • publication-ready quality — Publication-grade quality
  • clean line art — Clear lines
  • white background — White background
  • vector-like clarity — Vector-level clarity
  • labeled in [language] — Specify annotation language

Technique 2: Describe Layout Structure

ERNIE-Image excels at structured layouts. Explicitly describe panel count and arrangement:

  • multi-panel figure with four subfigures (A, B, C, D)
  • two-row comparison diagram
  • flowchart with six sequential steps
  • circular diagram with central concept and radiating elements

Technique 3: Use the PE Enhancer

For prompts containing extensive technical terminology, enable the Prompt Enhancer (PE):

  • PE expands brief technical descriptions into more detailed structured descriptions
  • LongTextBench with PE: 0.9733 vs. without PE: 0.9636
  • For academic scenarios, PE typically improves text annotation accuracy

Technique 4: Bilingual Chinese-English Annotations

ERNIE-Image supports both Chinese and English. For Chinese academic papers:

A scientific diagram showing [topic], 中文标注, academic style, white background,
clearly labeling each part: [Part A], [Part B], [Part C]

III. ComfyUI Batch Generation Workflow

For projects requiring batch generation of large numbers of scientific illustrations (series of papers, textbook illustrations), ComfyUI batch workflows are the optimal choice:

Workflow Architecture

  1. Prompt list loading: Read multiple prompts from a text file
  2. ERNIE-Image Base inference: Generate high-quality base images
  3. SUPIR upscaling (optional): Boost resolution to publication requirements
  4. Batch export: Uniform format output

Recommended Settings

Parameter Recommended Value Notes
Model ERNIE-Image Base High quality, suitable for academic illustrations
Steps 30-50 Academic illustrations need higher precision
CFG Scale 4.0 Maintain instruction-following capability
Resolution 1024×1024 or 1024×768 Choose based on chart aspect ratio
PE Enabled Improve text annotation accuracy
Sampler DPM++ 2M Balance quality and speed

IV. Comparison with Professional Tools

Tool Strengths Weaknesses Best Use Case
ERNIE-Image Free, open-source, strong text rendering Limited data precision Concept diagrams, mechanism figures, abstracts
Illustrae Purpose-built scientific illustrations Paid subscription Professional scientific illustrations
FigCanvas Research chart templates Paid Paper figures
R/ggplot2 Data precision Steep learning curve Real data visualization
Illustrator Full control Time-consuming, expensive Detailed hand-crafted illustrations

Key recommendation: ERNIE-Image is best suited for conceptual, schematic academic illustrations (mechanism diagrams, flowcharts, abstract figures). For figures requiring precise data display (experimental data charts), professional tools like R/ggplot2 or Python matplotlib should still be used. AI-generated data charts may have inaccurate numerical values.

V. Practical Case Studies

Case 1: Protein Signaling Pathway Figure

Requirement: Generate a teaching illustration of the EGFR signaling pathway

Prompt:

A professional scientific diagram showing the EGFR signaling pathway in cancer cells.
Top: EGFR receptor on cell membrane with ligand binding. Middle: downstream cascade
showing RAS, RAF, MEK, ERK pathway with phosphorylation arrows. Bottom:
nuclear translocation and gene expression. Clean academic style, labeled proteins,
color-coded compartments (membrane in blue, cytoplasm in light gray, nucleus in green),
white background, publication-quality illustration

Case 2: Machine Learning Workflow Comparison

Requirement: Compare traditional ML vs. deep learning workflow differences

Prompt:

A side-by-side comparison diagram. Left panel labeled "Traditional ML":
feature engineering → model training → prediction. Right panel labeled "Deep Learning":
raw data → automatic feature learning → prediction. Clean arrows, numbered steps,
academic illustration style, white background, clear English labels

VI. Caveats and Limitations

Things to Watch Out For

  1. Numerical accuracy: Numerical values in AI-generated data charts may be inaccurate; not recommended for displaying real experimental data
  2. Text length limits: Despite ERNIE-Image's strong text rendering, labels exceeding 50 words may still have errors
  3. Specialized symbols: Mathematical formulas, chemical equations, and highly specialized content may require post-processing correction
  4. Copyright considerations: Apache 2.0 license permits academic use, but image copyright attribution depends on specific journal requirements

Recommended Workflow

For high-quality academic illustrations, a AI generation + manual correction hybrid workflow is recommended:

  1. ERNIE-Image generates base illustration
  2. Inkscape (free) or Illustrator for post-correction (adjust text, fix incorrect labels)
  3. Export as high-resolution PNG/SVG
  4. Insert into paper typesetting software (LaTeX/Word)

Conclusion

ERNIE-Image, with its outstanding text rendering capability (LongTextBench 0.9733) and complex layout generation, is emerging as a new choice for researchers generating academic illustrations. It is particularly well-suited for schematic illustrations like concept diagrams, mechanism figures, flowcharts, and abstract figures.

The key advantages are: free, open-source, deployable locally, supporting bilingual Chinese-English annotations. For scenarios requiring precise data display, combine with professional data visualization tools.

Sources: ERNIE-Image GitHub repository, ERNIE-Image Technical Report (arXiv:2605.25347), Illustrae official documentation, FigCanvas official documentation

ERNIE-Image Team