How Many Labels Should a Scientific Diagram Have? A Practical Rule

Before-and-after redesign of an overloaded signaling cascade reduced to five essential steps

The cascade above once had two dozen labels. It now has five, and it teaches better — that's not an accident, it's the single most repeatable improvement in scientific figure design. This guide gives the label-count rule we use on every plate on this site, the reasoning, and the tiering system that lets you keep the other nineteen labels without wrecking the figure.

Apply it on the presets · Errors and fixes · Label rendering test

The rule

A scientific diagram carries at most six labels. Past six, split the topic or tier the labels.

Six is where three curves cross:

Label rendering. In our 20-plate test, quoted labels of 1–2 words render reliably as a set up to about six; beyond that, individual labels start dropping and spelling degrades — every added label is another roll of the dice, and the failures aren't independent (crowding compounds).
Leader-line geometry. Labels sit outside the figure. With ≤6 leader lines, the eye can trace each one to its target; past that, lines cross, and a crossed leader line is a wrong answer waiting to be read. There is no layout fix — the density is the problem.
Working memory. A figure is parsed in seconds. Readers hold a handful of items from one glance; a diagram whose labels exceed that stops being read and starts being searched. Search belongs in tables, not figures.

Why "more labels = more rigorous" feels true and isn't

Comprehensiveness feels scholarly. But the figure's job is communication, not storage — the complete enumeration belongs in the text or caption. The overloaded cascade at the top looked more rigorous and taught less: reviewers of the redesign don't miss the nineteen omitted labels; they can finally see the five that matter.

The counter-signal to watch for: if your figure needs its caption read aloud to be understood, it has too many labels. That's the practical test, and it works on other people's figures at a glance.

The fix that isn't deletion: tiering

You rarely have to lose the extra labels — you have to demote them:

Tier	Where it lives	Example
1. On-figure labels	On the plate, ≤6	TRENCH, SLAB, MAGMA, VOLCANIC ARC
2. Caption terms	In the caption, referenced by leader line number	"3: earthquake-depth dots (Wadati-Benioff zone)"
3. Companion table	In a key beside or below	every structure, spelled out
4. Text	In the prose	everything else

Tier 2 (numbered leader lines with a legend) is the professional default in journal figures — numbers are short, render perfectly, and never collide. It's also an AI-drafting trick: ask the model for numbered markers, then typeset the legend yourself where spelling is guaranteed.

Label design rules (the other half)

Count gets labels onto the figure correctly; design makes them readable:

1–2 words, all caps, quoted in the prompt. Long labels shrink to fit and stop being legible (why).
One side, one alignment. Labels on alternating sides of the figure halve usable line-length and double collisions. Outside, one column where possible.
Parallel leader lines. Lines that run together in the same direction read as a system; lines fanned from one point read as a star chart.
Labels never touch the drawing. A label crossing the illustration boundary is unreadable against structure.
Spell out symbols. ALPHA not α, CO2 not CO₂ — subscripts are the weak point of both rendering and reading.

The exceptions (when more is right)

Reference charts — an anatomical atlas plate or cloud-classification chart is a lookup figure, read with a finger, not parsed in a glance. Twenty labels is correct there. The tell: users of the figure expect to search it.
Numbered-marker figures — effectively unlimited labels because the on-figure payload is numbers only.
Assessment figures — a worksheet figure with three labels and five blanks is a test, not a teaching figure; different genre, different rules (teacher's guide).

The pattern in all three: the extra labels moved to a system — lookup, legend, or blanks — rather than free-floating text.

How this shows up in prompts

A [SUBJECT] figure with exactly six labels: [MAIN VIEW], one magnified inset,
labels arranged in a single column outside the figure with parallel leader lines.
Label "TERM 1", "TERM 2", "TERM 3", "TERM 4", "TERM 5", "TERM 6".
Title: "SHORT TITLE".

exactly six labels — the constraint belongs in the prompt, not in your hopes.
single column outside the figure with parallel leader lines — layout instruction that measurably reduces collisions.
Six filled, everything else demoted per the tier table before you generate — deciding tiers first is what prevents the two-dozen-label draft.

Scientific review checklist

[ ] Are all labels spelled correctly?
[ ] Six or fewer on-figure labels — or a deliberate exception genre?
[ ] Do leader lines reach the right structures without crossing?
[ ] Is the demoted information still available (caption, key, text)?
[ ] Are positions and proportions reasonable against your reference?
[ ] Has it been checked against a trusted source?

FAQ

Is six a hard limit?

Six is where three failure curves (rendering reliability, leader-line geometry, working memory) cross for on-figure text — treat it as a design speed limit, not a law of nature. Reference charts and atlas plates legitimately carry twenty labels because they're lookup figures, read with a finger. The question to ask isn't "can I fit more labels?" but "is this figure parsed in a glance or searched on demand?" — glances get six, searches get twenty.

Do numbered markers count toward the six?

The numbers themselves barely do — which is the trick. A figure with twelve numbered leader lines pointing to a typed legend carries twelve labels' worth of information with two words' worth of rendered text. That's why journals love numbered markers, and why prompting for numbers sidesteps the rendering reliability question entirely.

What about bilingual labels?

Halve the on-figure count. Two languages on one plate doubles the text density and the collision risk, so a figure that would carry six monolingual labels carries three plus a bilingual legend. Bilingual rendering is a specific model capability — GLM's image models treat Chinese-English text rendering as a headline strength — but layout physics still applies: two languages want a legend, not doubled leader lines.

How does this apply to assessment figures?

Assessment figures run different rules: fewer labels (three), plus blanks the student fills. A worksheet figure with five labeled structures and three blanked ones is a test instrument, not a teaching figure — see the teacher's guide for that genre's conventions. The six-label rule governs teaching and communication figures; tests deliberately withhold.

Building a numbered-marker figure

The tier-2 pattern in four steps: (1) prompt the figure with structures unlabeled, asking for "numbered circular markers 1 through 6 placed on each structure"; (2) verify the numbers sit on the right structures — this is now a placement check, much faster than a spelling check; (3) typeset the legend beside the figure, where your keyboard guarantees spelling and subscripts; (4) match numbers to legend in reading order (top-left clockwise reads naturally). You've moved the text burden from the model's weakest channel to your strongest one, and the figure survives translation, font changes, and journal reformatting without a regeneration.

Case study: the cascade, before and after

The signaling cascade that opens this page arrived with twenty-four nodes and crossing arrows — technically correct, functionally illegible, the figure equivalent of a paragraph with no sentences. The rebuild asked one question per label: does the reader need this term to get the finding? Five survived — receptor, the two intermediates that carry the story, the effector, the outcome. The other nineteen moved to a companion key.

A dense paragraph of methods text converted into a four-step visual schema Same science, five labels, and the figure now explains itself in a glance — which is why it leads this page instead of hiding in a supplement. When colleagues object that detail was lost, the honest answer is that it was moved: detail belongs where it can be searched, not where it can't be seen.