GLM-Image
Sign In
GLM-Image model weights are available on Hugging Face. Download
Tutorial

How to Create a Scientific Illustration with AI: Step-by-Step Workflow

A 4-step workflow for labeled scientific illustrations with AI: prompt anatomy, real examples, verification checklist.

1 min read
Sep 3, 2026
By Grace Bennett

Labeled botanical plate of a sunflower head with capitulum, disc florets, ray florets and parastichy spirals

This sunflower plate — with its labeled capitulum, disc florets, ray florets and counted spiral families — was created from a single prompt. The first draft did not require drawing or vector software. This guide shows the workflow behind it: how to pick a subject, write a prompt for the required labels, iterate on the result, and verify the science before you use it.

Generate your first illustration · Jump to the prompt template · More examples

A scientific illustration fails in two ways: the drawing is wrong, or the labels are unreadable. Pretty is optional; correct is not. So the workflow below is built around those two failure modes — the prompt exists to force a correct structure, and the last step exists to catch what the model got wrong.

What you need before you start

Three things, none of them artistic:

  1. A subject narrow enough to draw on one plate — "the water cycle" is a diagram, "how water moves through a leaf" is an illustration.
  2. A reference source you trust: a textbook page, a university lab guide, a journal figure. You will check your result against it in Step 4.
  3. A label list — the 3–6 structures that must be named on the final image. Write it before you write the prompt. The label list is the specification; the picture exists to point at the labels.

Step 1: Choose the subject and steal its vocabulary

Start from the reference, not from imagination. Open your source, list the structures it names, and use those exact terms — the scientific ones, not the casual ones. "Disc floret" instead of "the middle part"; "conduit" instead of "the tube."

For the sunflower plate above, the reference vocabulary was: capitulum, ray floret, disc floret, parastichy (the spiral arrangement). Those four terms decided everything downstream — the model drew one large head, one floret close-up, and a spiral overlay because the labels demanded places to point.

Subject Sunflower head (botany)
Model grok-imagine-image-2.0, generated on glmimage.app
Aspect ratio 16:9 (1792×1024)
Labels CAPITULUM, RAY FLORET, DISC FLORET, PARASTICHY
Worked well Clear three-panel layout; spiral overlay is countable; labels sit outside the drawing with room to read
Verify before use Parastichy counts should be Fibonacci-family numbers (21/34/55); check the ray/disc floret distinction is anatomically right for a sunflower rather than a daisy

The prompt behind it:

A mathematical-botany plate of the sunflower head: whole capitulum, a disc-floret
close-up, and an overlay spiral diagram showing Fibonacci parastichy families
counted and numbered in two directions. Label "CAPITULUM", "DISC FLORET",
"RAY FLORET", "PARASTICHY". Title: "COUNTING SPIRALS".

Notice what the prompt does: it names the panels, it fixes the label vocabulary, and it gives the plate a short title. You are not describing a mood; you are specifying a figure.

Step 2: Write the prompt

Every prompt in this guide uses the same five-slot template:

A [STYLE] plate of [SUBJECT]: [MAIN VIEW], [INSET OR SECOND PANEL],
[WHAT THE INSET SHOWS]. Label "LABEL 1", "LABEL 2", "LABEL 3",
"LABEL 4". Title: "SHORT TITLE".

Fill the slots like this:

Slot What it controls Example values
STYLE rendering style of the whole plate botanical plate, textbook diagram, museum infographic, cutaway, field-guide page
SUBJECT the organism, organ, or system animal cell, stratovolcano, bicycle drivetrain, DNA double helix
MAIN VIEW the primary orientation top view, cutaway, cross-section, Z-scheme flow
INSET the detail panel that earns its labels magnified organelle, close-up of florets, calculation panel
LABELS 3–6 terms, quoted, in caps keep them short enough to spell correctly at plate size
TITLE the caption printed on the plate 2–4 words, all caps

Three rules that matter more than the rest:

  • Quote the labels. Labels wrapped in quotes come out as text far more reliably than labels buried in a sentence.
  • Cap the label count at six. Every extra label is another chance for a spelling error and another line crowding the drawing. Six labeled structures done correctly beats twelve done approximately.
  • Ask for panels, not crowds. "One large cutaway plus one magnified inset" produces readable figures; "everything visible at once" produces soup.

Here is the same template applied to a cell-biology subject:

Cutaway animal cell with nucleus, mitochondrion, rough ER, Golgi and lysosome labeled

A large cutaway animal cell with its organelles arranged realistically, each
organelle shown once with a short functional caption beside it in a small grid
around the cell. Label "NUCLEUS", "MITOCHONDRION", "ROUGH ER", "GOLGI",
"LYSOSOME". Title: "THE ANIMAL CELL".
Subject Animal cell organelles (biology)
Model grok-imagine-image-2.0, generated on glmimage.app
Labels NUCLEUS, MITOCHONDRION, ROUGH ER, GOLGI, LYSOSOME
Worked well Five labels all rendered correctly; organelles in plausible positions
Verify before use Rough ER should sit near the nucleus; check the caption grid did not attach a caption to the wrong organelle

Step 3: Generate, then read the result like a reviewer

Generate, and before admiring it, interrogate it: Are all labels present? Spelled correctly? Pointing at the right structure? Is anything anatomically mirrored or invented?

An earth-science example, from prompt to result:

Stratovolcano cutaway from crater to magma chamber, labeled crater, conduit, dike, sill and magma chamber

A stratovolcano cutaway from crater to magma chamber: layered lava and ash
strata, central conduit, branching dike and horizontal sill, and convection
swirls within the chamber, eruption column above. Label "CRATER", "CONDUIT",
"DIKE", "SILL", "MAGMA CHAMBER". Title: "ANATOMY OF A VOLCANO".
Subject Stratovolcano internal structure (earth science)
Model grok-imagine-image-2.0, generated on glmimage.app
Labels CRATER, CONDUIT, DIKE, SILL, MAGMA CHAMBER
Worked well Dike vs. sill distinction came out correctly (dike cuts across strata, sill runs between them)
Verify before use Check the dike/sill orientation against the strata; confirm the eruption column matches the conduit geometry

When something is wrong, fix it in the prompt, by name: "the sill must be horizontal, between two strata layers" beats rerolling and hoping. One targeted correction per regeneration keeps the parts that already work. For a mechanical subject, the same loop applies — this bicycle drivetrain figure needed its gear-ratio inset specified explicitly to avoid a decorative-but-meaningless gear collage:

Bicycle drivetrain diagram with chainring, cassette, derailleur and jockey wheels labeled, plus gear-ratio inset

Step 4: Verify against your reference before using it

AI output is a draft until a human checks it. This is the step that separates a usable scientific illustration from a plausible-looking wrong one.

The image below is what verification looks like in practice: two generated heart cutaways side by side, the flawed one annotated with its actual errors — mirrored chambers and a stray vessel.

QA comparison of two AI-generated heart cutaways, one correct, one with circled errors

Run every result through this checklist:

Scientific review checklist

  • [ ] Are all labels spelled correctly?
  • [ ] Are all requested structures present?
  • [ ] Are positions and proportions reasonable against your reference?
  • [ ] Do arrows or flows point the right way?
  • [ ] Are stages shown in the correct order?
  • [ ] Has it been checked against a trusted source?

If any box fails, correct via the prompt loop in Step 3 or fix it by hand in a vector editor — then say so. For classroom use, a note like "AI-drafted, human-verified" is honest and increasingly expected; for publication or public science communication, treat the AI draft as the underdrawing and finish it properly. Our scientific drawing guide covers the hand-finishing path in more detail.

Make one prompt cover a dozen subjects

Because the template is slots, you can re-point it without rewriting it. Take the sunflower prompt and swap the subject line:

  • rose flower dissection — petals, stamens, pistil, receptacle
  • hibiscus flower anatomy — stigma, style, ovary, stamen column
  • daisy flower head — ray florets, disc florets, phyllaries
  • lotus flower cross-section — seed pod, receptacle, petal nodes

Same structure, new vocabulary from Step 1, same verification in Step 4. That is the entire method — which is why it scales across disciplines without you learning to draw each one. The botanical illustration guide and animal illustration guide apply it to their own kingdoms; the medical and anatomical guide applies it where the stakes — and the verification bar — are higher.

Five mistakes that ruin scientific illustrations

  1. Describing a vibe instead of a figure. "Beautiful science art of a cell" gets you wallpaper. Panels, views and labels get you a figure.
  2. Too many labels. Beyond six, spelling errors multiply and leader lines tangle.
  3. Trusting the first correct-looking output. Mirrored anatomy and invented structures are plausible — that is exactly why Step 4 exists.
  4. Letting the model invent the vocabulary. If the prompt does not fix the terms, the model will label the mitochondrion "powerhouse" — or something worse.
  5. Skipping the audience slot. A plate for a 7th-grade classroom wants bolder labels and fewer insets than a plate for a journal figure. Style follows audience.

FAQ

Can I use AI scientific illustrations in teaching materials?

Yes, as drafts or as verified figures — with the checklist above run honestly, and ideally with a disclosure. For graded exams or published curricula, verify every label personally; models occasionally produce confident, plausible, wrong anatomy.

How accurate are the labels, really?

Text rendering in current models is good for short, quoted, all-caps labels in English — that is exactly why the template enforces them. Accuracy of what the labels point to is a separate question and only Step 4 answers it. Never conflate the two.

Which disciplines work best?

Discrete structures with stable vocabulary work best: cells, organs, organisms, cutaways of machines and geological features. Highly quantitative figures (exact graphs, precise circuits) still need real tools — see our comparison in what is scientific illustration.

Do I own the output?

Generation on glmimage.app follows the site's standard content terms; scientific claims in the image are your responsibility either way. Attribute honestly when the science matters.

Start one now

Pick a subject you actually need, list its labels from a reference, and run the template. The scientific illustration page has discipline-specific presets for all nine areas — anatomy, botany, cell biology, engineering, astronomy and the rest — so the style and vocabulary slots are already filled in.

<!-- scientific-illustration-cluster-links:start -->

Continue through the scientific illustration collection

Create a scientific illustration · Browse the scientific illustration resource hub

Related: Scientific Illustration Prompt Refinement: Before and After Examples · From AI Draft to Editable Vector: GLMImage and Inkscape Workflow
<!-- scientific-illustration-cluster-links:end -->

<!-- scientific-authority-bridge:start -->

Choose a focused scientific illustration workflow

Use the broad scientific illustration generator when you are still choosing a discipline. For a defined deliverable, start with the focused workflow:

Whichever route you choose, keep the same sequence: define the brief, generate a draft, verify structures and labels, then revise before export.
<!-- scientific-authority-bridge:end -->

Tutorial Complete!

You've completed this tutorial. Need help? Check out our other tutorials or contact support.

How to Create a Scientific Illustration with AI: Step-by-Step Workflow | GLMImage Blog