sci-vis

A third of it is gone, or nine tenths

Hannah Isles — Mon, 12 Oct 2026 00:00:00 GMT

Somebody recording a tumour under a two-photon microscope has a decision to make at the end, and it is not the one about which frame to publish. It is what to say about everything that happened to the picture while they were recording it.

This file holds 47 frames of mouse tumour, each 21 optical sections deep, in three colours at once. Over those frames the picture empties out. The question this page is about is how much, and it has two defensible answers that are a long way apart.

One scale, all the way through

The first frame, and the scale every frame after it is drawn on. Three labels in one piece of mouse tumour: the transferred T cells on the left, the tumour cell nuclei in the middle, the blood vessels running through both on the right.

Frame 0 of 46.
Eight frames in. Very little has changed, which is worth showing: the display is doing nothing yet.

Frame 8 of 46.
Sixteen. The T cell panel is down to 60% of its signal while the vessels hold 76%.

Frame 16 of 46.
Twenty-four. All three are declining now, and at not very different speeds.

Frame 24 of 46.
Thirty-two. Still about half of each. Nothing so far looks like a cliff.

Frame 32 of 46.
Forty. Here is the cliff. Every channel has dropped to under a quarter, together, in eight frames.

Frame 40 of 46.
The last frame. 1% of the T cell signal is left, 4% of the tumour nuclei, 9% of the vessels.

Frame 46 of 46.

Seven of the forty-seven frames, on one scale

Left to right in every frame: T cells, tumour nuclei, vessels. Each channel is drawn on a single scale taken from its own first frame and held for the whole series, so what changes here is the signal rather than a display setting. The three colours are this site's, not the file's, whose stored colours disagree with its own channel names.

Source Weigelin B, Bolaños E, Teijeira A and others, Proceedings of the National Academy of Sciences, 2015. CC BY 4.0. CC BY 4.0

Data IDR idr0026, image 3261701 · retrieved 2026-10-09 · CC BY 4.0
ProcessingMaximum across all 21 planes at each frame, read from the deposited OME-Zarr by scripts/pieces/fading-rates.mjs. Nothing resampled.

Almost everything goes, and it mostly goes at the end rather than steadily. That much the pictures settle. Putting a number on it is where it gets interesting.

Two numbers for the same pixels

The same three channels, measured two ways

On the left, each channel's 99th percentile as a fraction of its own first frame. On the right, the same percentile with that frame's own median subtracted first, the median of a max projection of this tissue being background. Same pixels, same frames, one subtraction between them. The two panels share a y axis, which is the only reason they can be compared at all.

Source Weigelin B, Bolaños E, Teijeira A and others, Proceedings of the National Academy of Sciences, 2015. CC BY 4.0. CC BY 4.0

Fraction of frame 0 remaining at frame 46, both ways, from `data/idr0026-ctl/fade.json`.
Channel	as recorded	above background	Baseline at frame 0
T cells	18%	1%	1589
tumour nuclei	29%	4%	2399
vessels	69%	9%	342

Read as recorded, the vessels lost 31% and the T cells 82%, and the obvious conclusion is that these labels behave very differently. Subtract the background first and the vessels lost 91% and the T cells 99%, and the obvious conclusion is that they behave much the same.

The left panel is not wrong about anything. It is a true statement about the numbers in the file. It is simply not a statement about the dye, because every pixel in a channel carries that channel’s baseline whether or not anything is glowing, and the vessel channel’s baseline is most of what it has. Its 99th percentile at the first frame is only about 144 above its own median. A collapse of the signal cannot move a ratio built on that by very much, so it does not, and the number comes out reassuring.

This site measured it the first way and nearly published the reassuring version. What caught it was not a review of the arithmetic. It was putting the vessel row next to the claim in a poster and seeing a strip of black squares under the words “31 per cent gone”.

The other way the same frames can mislead you

The T cell channel at frame 0 and frame 40

Top row: both frames on the scale set by the first, which is what every figure above uses. Bottom row: each frame scaled to its own range, which is the default in most software. The pixels are identical between the rows. Only the mapping from value to brightness differs.

Source Weigelin B, Bolaños E, Teijeira A and others, Proceedings of the National Academy of Sciences, 2015. CC BY 4.0. CC BY 4.0

The bottom right panel is built from pixels holding a small fraction of the signal of the panel above it, and it does not look like it. Autoscaling is not a trick and nobody switches it on in order to mislead. It is the default, it is usually what you want while you are hunting for something, and it removes the one property that would have let a reader compare two frames.

So there are two subtractions in this piece, and they are the same mistake in different clothes. Scaling each frame to itself takes away the comparison between frames. Not taking the baseline off takes away the comparison between channels. Both leave a number or a picture that looks fine.

What this is not an argument for

Not against the recording, which does something no single exposure can, and which was deposited whole so that this page could be written by somebody who has never held a two-photon microscope.

Not against these authors. Their figures are not reproduced here and are not the subject. This is one image from the control arm of their experiment, the annotation recording it as cytotoxic T lymphocytes only, and nothing here speaks to the question their paper was asking.

What is not measured: why it fades. Bleaching is the obvious guess and not the only one. The vessel label is a dextran injected into the bloodstream, which leaves it over time, while the other two labels are carried by the cells themselves. Three curves may be falling for three reasons, and the sharp drop near the end arrives in all three at once, which is the sort of thing that happens to a preparation rather than to a fluorophore. Which of those it was is a question for whoever ran the experiment.

And what the file does not say: how long. The deposit records no interval between frames and no pixel size. Everything here is counted in frames, and no figure on this page carries a scale bar.

This image has twenty-seven channels and your figure has three

Hannah Isles — Fri, 09 Oct 2026 00:00:00 GMT

A fluorescence microscope runs out of colours at about four. The dyes overlap, the filters bleed, and past that nothing can be told from anything else. So a panel is four markers, and a question needing a fifth needs a second experiment.

Imaging mass cytometry does not work that way. Each antibody carries a different metal isotope, and the instrument counts the metals by mass, so the channels cannot bleed into one another at all. This section of human tonsil was recorded with 27.

The page is still flat and still has three colours.

The nuclear counterstain. Every cell in the section, which is what makes it the right thing to measure the others against: a marker covering a tenth of the section and a marker covering a tenth of the cells are different claims.

DNA1-191/193Ir, marking 86.6% of the tissue.
CD19, a B cell marker. The follicles appear, because that is what a follicle is made of.

CD19-169Tm, marking 8.6% of the tissue.
CD3, a T cell marker. It lights the spaces between the follicles, almost exactly where the last panel was dark. Two channels, two populations, one section.

CD3-170Er, marking 16.1% of the tissue.
CD68, found on macrophages. Not a region this time but a scatter of individual cells, present in both compartments and belonging to neither.

CD68-171Yb, marking 5.4% of the tissue.
And the blank. No antibody, carried through the run anyway. This is what the instrument records when there is nothing there, and it is the only reason to believe the panels above it.

empty, marking 1.9% of the tissue.

Five of the twenty-seven, one at a time

Each panel is one antibody, shown on its own. The last is the panel's own blank: an empty channel carried through the run, which is what the instrument records where there is nothing to find. It marks 1.86% of the tissue, and it is the thing to check any claim about the other channels against.

Source Durand M, Walter T, Pirnay T, Naessens T, Gueguen P, Goudot C, Lameiras S, Chang Q, Talaei N, Ornatsky O, Vassilevskaia T, Baulande S, Amigorena S, Segura E, Journal of Experimental Medicine, 2019. CC BY 4.0. CC BY 4.0

Data IDR idr0054, image 5025551 · retrieved 2026-10-07 · CC BY 4.0
ProcessingRead from the deposited OME-Zarr at pyramid level 2 by scripts/pieces/twenty-seven-channels.mjs. Nothing resampled.

Those four markers divide the section in ways that have nothing to do with one another. B cell regions, T cell regions, a population scattered through both. A figure showing any one of them is true and is not the tissue.

What three colours hold

Three channels of the twenty-seven

DNA, CD3 and CD19 laid over one another. We chose these three and the colours are this site's, not the archive's: it is the view this tissue is usually shown in, and it is a good figure. It is also three of the 27 channels that were recorded.

Data IDR idr0054, image 5025551 · retrieved 2026-10-07 · CC BY 4.0
ProcessingRead from the deposited OME-Zarr at pyramid level 2 by scripts/pieces/twenty-seven-channels.mjs. Nothing resampled.

That is a good figure and this piece is not an argument against it. The question is what it costs, and that can be counted.

Setting aside the two nuclear channels, which mark every cell by design, and the blank, there are 24 markers. Between them they mark 85.1% of the tissue in the section.

Take the three markers that between them cover the most, found greedily rather than chosen. They are CD185, CD1c, CD370, and they cover 50.2% of everything the 24 markers mark.

That is a ceiling and not an estimate. Any three picked for being informative rather than for being large cover less than that, including the three in the figure above. Half, at best, of what the instrument was asked to find.

It takes 11 channels to reach 90%. Not three, not four, and not a number a page can carry.

All twenty-seven

Every channel in the file, in the order it is stored, each on its own and each scaled to its own range. The blank is the dark one in the second row. Measured at pyramid level 2 of 5, which is 675 by 675 of the 2700 by 2702 recorded.

Data IDR idr0054, image 5025551 · retrieved 2026-10-07 · CC BY 4.0
ProcessingRead from the deposited OME-Zarr at pyramid level 2 by scripts/pieces/twenty-seven-channels.mjs. Nothing resampled.

Every channel and how much of the tissue it marks, from `data/idr0054-tonsil/channels.json`. Each is cut on its own histogram by Otsu.
Channel	Share of tissue marked
DNA2-191/193Ir (counterstain)	90.61%
DNA1-191/193Ir (counterstain)	86.63%
CD185/CXCR5-151Eu	18.99%
CD1c/BDCA1-biotin + Neutravidin-173Yb	17.60%
CD45RA-155Gd	17.46%
CD370/Clec9A-161Dy	16.22%
CD3-170Er	16.08%
CD11b-149Sm	15.63%
HLA-DR-174Yb	15.03%
Bcl6-163Dy	14.36%
CD14-156Gd	12.93%
CD45-152Sm	12.71%
IL-21-164Dy	11.81%
ICOS-148Nd	11.03%
CD123-143Nd	10.55%
CD19-169Tm	8.58%
CD141/BDCA3-165Ho	7.67%
CXCL13-157Gd	7.47%
CD279/PD-1-175Lu	7.28%
CD324/E-Cadherin-158Gd	6.23%
CD303/BDCA2-147Sm	6.06%
CD56-176Yb	6.01%
CD206-168Er	5.81%
CD68-171Yb	5.36%
CD11c-159Tb	4.74%
empty (blank)	1.86%
alphaSMA-141Pr	1.12%

The blank is the reason to believe any of this

Every number here rests on deciding, per channel, which pixels hold signal. That is a threshold, and a threshold can be moved until the answer is the one you wanted.

The panel carries its own control. One channel had no antibody in it and was carried through the run regardless, and whatever method is used to say how much a channel found has to say that this one found almost nothing. It ranks 26 of 27 and marks 1.86% of the tissue. The script that produced these figures refuses to write them if that stops being true.

That is not a flourish. It is the difference between a measurement and a number, and it was in the authors’ data rather than ours: they put the blank in the panel because they intended somebody to check.

What to do with twenty-seven channels and three colours

Nothing here says publish a twenty-seven panel figure. Nobody would read it, and the grid above is legible only because it is making an argument about itself rather than about tonsils.

What is worth saying is smaller. The figure is a selection, the selection is an argument, and the reader cannot see which twenty-four things were set aside. Saying how many channels were recorded costs one clause in a legend. Depositing the rest, which these authors did, costs nothing at all and is why this page could be written by somebody who has never touched the tissue.

The same number twice, and two different cells

Hannah Isles — Tue, 06 Oct 2026 00:00:00 GMT

A human cell divides in about an hour. A figure of it is one frame.

This one was imaged every 90 seconds for 60 minutes by confocal microscopy, triggered automatically when the cell entered prophase, and the whole stack was deposited. So it is possible to ask what any single frame of it would have left out.

The cell has just entered prophase. The chromatin is spread through the nucleus and there is little that is bright.

Frame 0 of 40, 0 minutes in. 280 pixels above the cut.
Fifteen minutes in. The chromosomes are condensing: the same DNA, packed into less space, so each pixel holding it is brighter.

Frame 10 of 40, 15 minutes in. 487 pixels above the cut.
Twenty-seven minutes. This is as condensed as it gets, and it is the frame a figure would almost certainly use.

Frame 18 of 40, 27 minutes in. 665 pixels above the cut.
Forty-three minutes. Two separate masses, pulling apart. The amount of condensed chromatin is already falling.

Frame 29 of 40, 43.5 minutes in. 134 pixels above the cut.
Fifty-eight minutes. Two daughter nuclei, decondensed, and almost nothing above the threshold any more.

Frame 39 of 40, 58.5 minutes in. 64 pixels above the cut.

One cell, one hour, five of forty frames

The DNA channel, projected over all 31 optical sections at each timepoint. Frames are taken at even positions through the series rather than chosen for what they show, so the sequence is sampled rather than curated. The other nuclei in the field belong to cells that are not dividing and sit still for the hour.

Source Walther N, Hossain MJ, Politi AZ, Koch B, Kueblbeck M, Ødegård-Fougner Ø, Lampe M, Ellenberg J, Journal of Cell Biology, 2018. CC BY 4.0. CC BY 4.0

Data IDR idr0052, image 5514375 · retrieved 2026-10-06 · CC BY 4.0
ProcessingDNA channel, maximum over all 31 planes at each timepoint, read from the deposited OME-Zarr at pyramid level 0 by scripts/pieces/mitosis-time.mjs.

What rises and falls

Condensed chromatin is brighter per pixel than the same DNA spread through an interphase nucleus, so counting the pixels above a fixed intensity follows condensation without anything having to be outlined or identified.

It climbs to a peak at 27 minutes and falls away. That peak sits in the same place at every cut tried: 28.5 min, 27 min, 27 min at the 99.0th, 99.5th, 99.9th percentiles respectively. The height of the curve changes with the cut. Where the top of it falls does not.

The two frames this piece is about

7.5 min, on the way in · 415 px

31.5 min, on the way out · 417 px

7.5 minutes and 31.5 minutes

Both frames hold the same amount of condensed chromatin: 415 pixels above the cut on the left and 417 on the right, a difference of 2. They are 24 minutes apart and the cell is doing the opposite thing in each. The measurement cannot separate them. Nobody looking at them needs it to.

Source Walther N, Hossain MJ, Politi AZ, Koch B, Kueblbeck M, Ødegård-Fougner Ø, Lampe M, Ellenberg J, Journal of Cell Biology, 2018. CC BY 4.0. CC BY 4.0

The same measurement at three cuts, from `data/idr0052-mitosis/mitosis.json`. The peak does not move.
Cut	Intensity	Peak	First frame	At the peak	Last frame
99.0th percentile	23369	28.5 min	628 px	990 px	260 px
99.5th percentile	27380	27 min	280 px	665 px	64 px
99.9th percentile	36190	27 min	16 px	255 px	2 px

A quantity that is the same in both. A cell that is doing opposite things in each. The number is not wrong, and it is not even imprecise: those two frames really do hold the same amount of condensed chromatin. It is simply not a quantity that distinguishes going in from coming out, and no amount of decimal places would make it one.

The pictures separate them in about a second, and they do it without anybody deciding in advance what to measure. That is the thing a visualisation is for which a table is not: it carries the properties nobody thought to ask about.

One caution about those two frames. The site measured a second quantity, the number of separate bright regions in each, and it comes out at 48 on the left and 23 on the right. That is a real difference and it is not a count of chromosomes or of chromosome masses: at these thresholds the chromatin resolves into many bright fragments rather than two tidy bodies. It is reported because it differs, not because it names anything.

All forty frames

The whole series at once, in order, left to right and top to bottom. Each tile is one timepoint, ninety seconds apart. This is the figure that cannot be made from a single exposure, and it is the one that makes the direction obvious.

Source Walther N, Hossain MJ, Politi AZ, Koch B, Kueblbeck M, Ødegård-Fougner Ø, Lampe M, Ellenberg J, Journal of Cell Biology, 2018. CC BY 4.0. CC BY 4.0

What this is not an argument for

Not against measuring. The curve above is a measurement, it is the thing that found the two frames, and no amount of looking at forty tiles would have turned up a pair matching to two pixels.

It is an argument against the order people usually put them in. The number is not the finding with the picture as decoration. Here the number raises the question and cannot answer it, and the picture answers it immediately. Both were in the same deposited file, free to anyone, for eight years.

A third of this image is hiding something behind it

Hannah Isles — Tue, 06 Oct 2026 00:00:00 GMT

A confocal microscope does not take a photograph. It takes a stack of optical sections, each one a thin slice of the specimen at a known depth, and the figure in the paper is almost always a single image made by flattening them.

This stack is 236 sections spanning 118 µm. The usual way to flatten it is a maximum-intensity projection: for every position in the image, look along the line of sight through all 236 sections and keep the brightest value. It is the right default, it is what everybody does, and it is in every imaging package as one click.

It is also a decision about what to do when two things are behind one another, and the decision is to show one of them.

How often that happens here

Along each line of sight, this counts the separate runs of signal: stretches of sections holding something, divided by stretches holding nothing. Where a line of sight crosses two separated runs, the projection has two objects to choose between and returns one.

Of the 29,949 positions in this image where there is any signal at all, 10,032, or 33.5%, hold two or more separate objects stacked behind one another. Where that happens the two are typically 33 µm apart, which is more than a quarter of the depth of the whole stack.

The projection

Where it fuses two or more objects

The projection, and where it is hiding something

Left: the maximum-intensity projection of the DAPI channel, which is the image a paper would carry. Right: the same picture with every position marked where the line of sight crosses two or more separate objects. Both are built from the same array in the same script, so the picture and the number cannot disagree. The marking is not a correction; the point is that the left image looks exactly as trustworthy in those places as everywhere else.

Source Blin G, Sadurska D, Migueles RP, Chen N, Watson JA, Lowell S, PLoS Biology, 2019. CC BY 4.0. CC BY 4.0

Data IDR idr0062, image 6001240 · retrieved 2026-10-06 · CC BY 4.0
ProcessingRead from the deposited OME-Zarr at pyramid level 0, DAPI channel, all 236 sections, nothing resampled. Measured by scripts/pieces/projection-loss.mjs.

The same measurement at thresholds chosen before the result was known, from `data/idr0062-nucleus/projection-loss.json`.
Threshold	Value	Positions fusing two or more	Median separation
Otsu	342	33.5%	33 µm
90.0th percentile	191	38.6%	38 µm
95.0th percentile	450	34.3%	24.5 µm
97.5th percentile	673	31.0%	16 µm
99.0th percentile	888	22.4%	9 µm

Why the threshold is in that table

Deciding that a voxel holds signal needs a threshold, and a threshold can be chosen to produce a number. So this one is not chosen: it comes from Otsu’s method, which takes it from the histogram of the stack. The table then repeats the whole measurement at four fixed percentiles either side of it.

The answer moves, from 38.6% at the loosest cut to 22.4% at the strictest. It does not stop being true. Even at the threshold that throws away all but the brightest one per cent of the stack, a fifth of the remaining image is still stacking separate objects on top of each other.

What the measurement does not say is that both objects are nuclei. This counts runs of signal along a line of sight; it does not segment anything and it does not identify what it found. Two separated runs 33 µm apart are very unlikely to be one object, and whether the second one is a nucleus, a fragment or something else is a question for whoever was at the microscope.

Seeing them come apart

An interactive view of the same confocal stack, one optical section at a time. In a single section about ten nuclei are in focus, each showing LaminB1 as a bright ring around a darker interior. Moving through the sections, nuclei appear and disappear as the focal plane passes through them, so objects that overlap in the projection are plainly at different depths. The projection view shows roughly twenty-five nuclei at once with every ring filled in.

Every optical section in the stack, in order. Move through them one at a time with the slider once it appears.

The same specimen, one optical section at a time

Move through the 62 sections and the nuclei separate: at any one depth about a third of them are present and the rest are elsewhere. Switch to the projection and they return to one plane, which is the image the measurement above is about.

Source Blin G, Sadurska D, Migueles RP, Chen N, Watson JA, Lowell S, PLoS Biology, 2019. CC BY 4.0. CC BY 4.0

What this is not an argument for

Not for abandoning projections. A projection is how you show twenty-five nuclei at once, and the alternative on paper is twenty-five panels nobody will look at. The authors of this study deposited the whole stack, which is the thing that made any of this checkable, and that is the standard worth copying.

It is an argument about what a projection is. Not a summary of the stack, and not a photograph of the specimen: a specific reduction that resolves every collision silently and in the same direction. What it discards is not dimmed or blurred at the edges. In 33.5% of the image, something is simply not there, and nothing in the picture says which part.

The stack is deposited and openly licensed. If a figure matters enough to argue about, the sections behind it can be opened and looked at, which is a thing that was not true of published microscopy twenty years ago and is quietly true of a great deal of it now.

What the sentence about an antibody leaves out

Hannah Isles — Mon, 05 Oct 2026 00:00:00 GMT

Anyone who has taken an immunology course can draw an antibody. A Y. Two arms that grip whatever the antibody is against, one stem that tells the rest of the immune system what to do about it. The sentence is true, the drawing is useful, and between them they have taught the shape to everybody who knows it.

They also agree with each other rather than with the molecule. One intact antibody has been sitting in the Protein Data Bank since 1997, deposited by Harris and colleagues at 2.8 Å, free for anyone to download. It is possible simply to measure it.

18 sugar residues

This is the antibody everyone draws: two arms of the same length, placed symmetrically about the stem. It is the picture that goes with the sentence, and the sentence does not contradict it.

A symmetrical Y, both arms the same length from the stem.
This is where the three lobes actually sit in the deposited structure, to the same scale. The arms are 75 degrees apart, and they do not straddle the stem evenly.

The arms move to their measured positions, 75 degrees apart and no longer mirrored.
The two arms are not the same distance from the stem either: 80.3 Å on one side and 72.2 Å on the other. The whole molecule spans 16.3 nm.

Each arm is labelled with its measured distance from the stem.
And it is sugared. The structure holds 18 sugar residues, 220 atoms in all, and every one of them is on the stem. None is on the arms.

The stem is recoloured to mark it as the site of all the sugar in the structure.

The drawn antibody and the measured one

Centres of mass of the three lobes, to scale, in the plane the three of them define. The lobes are found by clustering the deposited atom positions, not by cutting the chain at a residue number: the split is confirmed by the stem containing 0 of 3294 light-chain atoms, light chains occurring only in the arms.

Source Harris LJ, Larson SB, Hasel KW, McPherson A, 1997. CC0 1.0. CC0 1.0

Three numbers, none of which a sentence carries well. The arms sit 75 degrees apart. They are 80.3 and 72.2 Å from the stem, which is a difference of about a tenth. And all 220 sugar atoms in the structure are on the stem, with none on the arms.

That last one is not decoration. It is most of what distinguishes this part of the molecule from a drawing of it, and a Y with two arms and a stem has nowhere to put it.

Which raises the question the drawing cannot be asked. A Y on a page has two dimensions. How many does the molecule use?

The honest difficulty, which is also the point

A crystal structure is one molecule, held still, in one pose. The joint between the arms and the stem is flexible, which is the whole reason an antibody can grip two things at once that are not conveniently spaced. So 75 degrees is what this molecule was doing in this crystal. It is not the angle of antibodies, and anybody quoting it as one would be doing with a number exactly what the Y does with a shape.

Which is the argument for turning it round rather than looking at a picture of it. A still image, ours included, has to choose a viewpoint, and every viewpoint flattens one of the three dimensions the molecule is using. The four frames above are honest about the measurement and still show you a projection.

A three-dimensional view of an intact antibody. Two arms extend from a central stem at an angle to one another, and the whole molecule is clearly not flat: rotating it shows the arms standing out of the plane of the stem rather than lying in it.

The still above shows the same structure. Loading the viewer lets you rotate and zoom it.

The deposited structure

PDB 1IGT, an intact IgG2a monoclonal antibody at 2.8 Å, served from this site rather than fetched from RCSB so the page does not depend on somebody else's uptime. Nothing is modified: this is the file as deposited.

Source Harris LJ, Larson SB, Hasel KW, McPherson A, 1997. CC0 1.0. CC0 1.0

Turn it, and the thing the diagram cannot say says itself. Along its own three principal axes the molecule measures 163 by 125.6 by 69.3 Å. That third number is 43% of the first, and it is the one a flat drawing has nowhere to put. Not a detail lost at the edges: a dimension, nearly half the size of the longest one, gone.

This is what a visualisation is for. Not to decorate the explanation, and not to catch somebody out, but to carry the part of the finding that does not survive being put into words. The sentence about the Y is not wrong. It is just the part of the molecule that happens to fit in a sentence.

Both ends of this protein are the same colour

Hannah Isles — Sat, 19 Sep 2026 00:00:00 GMT

Structural biologists colour a protein chain from blue at one end to red at the other. It is not a stylistic tic. It is a documented mode in the standard tools: ChimeraX ships rainbow as a synonym for color sequential, which gives a different color for each residue in a biopolymer chain. Everyone in the field reads it without thinking, which is the definition of a convention working.

The usual complaint about rainbow colourmaps does not apply here. A rainbow is normally wrong because it invents order in a variable that has none, and residue number has order: the first residue really does come before the last. A sequential ramp is the right family of answer. The problem is which one.

Figure 3 from the RoseTTAFold paper

Six panels of protein structure, each coloured along the chain from one terminus to the other. Reproduced under CC BY 4.0 from the author accepted manuscript deposited in Europe PMC. That is deliberately not the typeset version in Science, which carries the Science Journals Default License and could not be reproduced here: the CC BY applies to the accepted manuscript, through the rights retention the paper's Wellcome funding requires.

The ramp runs blue, cyan, green, yellow, orange, red along the chain. Measured against matplotlib’s own definition of jet, this figure covers 52% of that locus contiguously, which is what separates a ramp from a set of categorical colours that happen to span the hue wheel.
On a surface rendering the ramp reads as territory. The eye takes the cyan-to-green edge for a boundary between two regions, when it marks nothing except that the chain has kept going.
The magenta and red spheres mark particular residues. They are a categorical mark laid over a continuous ramp that is already using the whole hue wheel, so the marker has no hue left that means only "look here".

Source Baek M and others, 2021. CC BY 4.0. CC BY 4.0

A sequential ramp does its work through lightness. That is what survives a monochrome printer, a bad projector and every kind of colour blindness, and what lets a reader put two points in order without consulting a key.

This ramp’s lightness does not run in one direction. Across the sixteen steps sampled in data/rosettafold it climbs and falls again, a cumulative reversal of 101.7% of its own range with a single drop of 38.5%.

As published

Desaturated

One panel, as printed in colour and in grey

Our derivative, not the authors' figure: panel C desaturated with the luminance weights this site's contrast check uses. The two ends of the chain differ by 0.008 in relative luminance. The middle of the chain differs from both by about twenty times that.

Source sci-vis.com, after Baek and others 2021, 2026. CC BY 4.0. CC BY 4.0

Relative luminance of the ramp at the two ends of the chain and at its midpoint, from `data/rosettafold/measurements.json`.
Position on the chain	Colour	Relative luminance
First residue	#0000bf	0.038
Middle of the chain	#ffff00	0.928
Last residue	#800000	0.046

Both ends of the chain arrive at the same tone. The middle, no more important than anywhere else, is the brightest thing in the picture. A ramp whose job is to say this comes before that has made the two ends indistinguishable and put the emphasis where nothing is happening.

None of which makes the structures wrong or the prediction less remarkable. It would have cost one argument to avoid: a perceptually uniform ramp carries the same ordering in lightness, so the chain still runs one way when the colour is gone. The convention is doing something real. It is just doing it with the one part of colour that does not survive being printed.

The figure with no text in it

Hannah Isles — Fri, 18 Sep 2026 00:00:00 GMT

Open the SVG your plotting script just wrote and search it for one of your own axis labels. On most people’s figures you will find the words, and you will not find any text.

That is not a riddle. Somewhere between your plt.savefig and the file on disk, each letter stopped being a letter and became a small drawing of one. The picture is pixel-for-pixel identical. Everything you would want to ask about the type afterwards, and everything your journal’s production team would want to ask, has quietly stopped having an answer.

What the two files look like

Here is one ordinary figure: 342 penguins, flipper length against body mass, three species, drawn in matplotlib 3.11.2 with nothing customised. It was exported twice from the same code with one setting changed, svg.fonttype, and nothing else touched.

The two images are indistinguishable. The two files are not.

What is inside the two exports

The same figure, exported twice. The setting named path is what matplotlib ships, and at that setting the file holds no text at all: every letter is a reference to an outline in the file's own <defs>, placed 109 times. Set it to none and the 20 pieces of text come back as text. Neither export looks different on the page.

Source sci-vis.com, 2026. Original work for sci-vis.com.

Measured from the two files by `scripts/pieces/outlined-type.py`.
Measurement	Default (`path`)	`none`
Text elements	0	20
Glyph outlines placed	109	0
Type sizes the file declares	none	10 pt
Fonts the file names	none	DejaVu Sans, and its fallbacks
Where the words “Flipper length (mm)” appear	in a comment only	in the text
File size	84,229 bytes	65,556 bytes

The last row of that table is what makes this hard to notice. The default export does still contain the string Flipper length (mm). matplotlib writes it into an XML comment beside the outlines it drew it as, as a courtesy. So if you go looking for your axis label in the file, you will find it, and conclude that everything is fine.

It is not there as text. It is there as a note explaining what the drawing is of. The check most people would run is the one check that passes either way.

It is not a matplotlib problem

We went looking for figures drawn by tools that are not installed here, on Wikimedia Commons, under licences that allow it. The answer was the same from two of the three, by two different mechanisms neither of which announces itself.

Measured by `scripts/pieces/outlined-type-survey.mjs` and `outlined-type.py`. Every attribution is sourced in `data/outlined-type/survey.json`.
Tool	Live type?	How the letters got there
matplotlib 3.11.2, as shipped	No	each glyph defined once in <defs> and placed with <use>, 109 times, from DejaVuSans
matplotlib 3.11.2, svg.fonttype: none	Yes	live type
Julia, Plots.jl	No	each glyph drawn as a <path>
R, the built-in svg() device	No	each glyph defined once as a <symbol> and placed with <use>, which is cairo’s way
gnuplot 4.2	Yes	live type

gnuplot is the control, and it is the one that behaves: it writes real text and its labels can be read straight out of the file. Julia’s Plots draws each letter as a path. R’s built-in svg() device is cairo-backed, so it defines each letter once and stamps it wherever it is needed. matplotlib works the same way as R, defining its glyphs as paths rather than as symbols. Different routes, same outcome, and no warning from any of them.

A caution about that table, since this site has now made the mistake twice. The Julia file is filed on Commons under Category: Created with Matplotlib and is nothing of the kind: the author published the Julia source on the same page. Every attribution above comes from what the author wrote beside the file or from a string inside the file, never from a category. A label applied by a third party is not evidence about what produced a file.

What it costs

Nothing, most of the time. The figure renders correctly, which is exactly why this survives all the way to submission.

It costs something at the two moments that matter. The first is when somebody asks a question about the type, and here is what that looks like on the very figure above.

The type in that figure, drawn and printed

The figure declares itself 132.1 mm wide and sets its type at 10 pt. Reproduced in a 85 mm single column, every size in it is multiplied by 0.643, and that type arrives at 6.4 pt against a floor of 8. The sizes below are shown in proportion to one another; how big they are on your screen depends on your screen, which is the whole difficulty with judging type by eye.

Source sci-vis.com, 2026. Original work for sci-vis.com.

That figure is too small at a single column, and it is too small in both exports, because the outlining changed nothing about the picture. The difference is that one of the two files can be asked and the other cannot. In the default export there is no type size to read, no font to name and nothing for a checker, a copyeditor or you in six months to measure. The problem is still there. The evidence of it has gone.

The second moment is when someone needs to change a word. A typo in a panel label, a gene name the journal wants italicised, a unit that changed in revision: all of these are a two-second edit in a file with text in it, and a redraw-from-the-script in a file without. If the script still runs. If you still have it.

The journal’s own account of this

PLOS says it plainly on its figure guidelines page, under Text within EPS figures, retrieved 2026-09-18:

Embed fonts, or convert them to outlines to prevent missing or improperly rendered text. In those files that are created in software like Matlab, open Illustrator or Inkscape and convert text to outlines. You will not be able to change your text after it has been converted to outlines.

Read that carefully, because it is not on our side of the argument. PLOS recommends outlining for EPS figures, and gives a good reason: a font that is not embedded goes missing in production and the text renders wrong or not at all. Outlining makes a figure immune to that. It is a legitimate and sometimes correct thing to do.

The same page also requires, of every figure, arial, times, or symbol font only in 8-12 point.

Both of those are PLOS’s, and together they describe the trap exactly. Convert your text to outlines and the second requirement becomes uncheckable, by you and by them. The figure may well comply. Nobody can demonstrate that it does.

The difference between PLOS’s advice and your tool’s default is when. Outlining is a finishing step, done deliberately, to a figure that is already the right size with the right type in it, on the way out the door. Your tool does it first, silently, to every figure, including the seventeen drafts you are still revising.

What to do about it

Set the export to keep your type as type, once, in whatever you plot with. For matplotlib that is svg.fonttype: none, in a matplotlibrc or with plt.rcParams at the top of the script. In R it is a different device rather than a setting: svglite::svglite() in place of the built-in svg().
Know what that costs, because it is a trade and not a free win. matplotlib’s own configuration file describes none as “assume fonts are installed on the machine where the SVG will be viewed”, and svglite says the same thing from the other side: outlining “does not require the font to be present on the system that displays the SVG but makes it more or less impossible to edit the text after the fact”. Live type is a file that needs its font; outlined type is a file nobody can correct. Choose per stage, not once and for ever.
Check what you actually have. Open the file and search it for <text, not for your axis label. Your label is in there either way; <text is only in there when the type is real. The figure checker on this site does the same thing and will tell you which kind you have.
Outline last, if at all, and only on the copy you are submitting. Keep the version with text in it, because that is the one you will edit at revision.
If you are sending EPS or PDF to PLOS, follow PLOS, not us: embed the fonts or outline them, as they ask. Do it to a figure whose type you have already checked.

One thing that is not a problem: this is a vector-export setting and it does not touch your PDF. The same figure saved as a PDF keeps its text either way, which is worth knowing when a journal takes both.

The one-page card

Keep your type as type (PDF, A4). The export setting for matplotlib and for R, each quoted from that tool’s own documentation, the line to paste, and the check to run on a file you already have. Tools whose setting we could not verify are named on the card as unverified rather than guessed at. Also on the site as a page.

SEM or SD?

Hannah Isles — Sun, 13 Sep 2026 00:00:00 GMT

It gets asked as though it were a preference, usually late, usually by someone who has already made the figure. Reviewers have opinions. Supervisors have habits. Neither is the answer, because the two are not two ways of drawing the same thing. They are two different quantities, and one of them is under your control in a way the other is not.

The arithmetic, first

Standard deviation describes how spread out your measurements are. Standard error of the mean describes how precisely you have pinned down their average. They are related by one line:

SEM = SD ÷ √n

Which means the error bar you draw depends on how many animals you used, and nothing about that is a claim about the animals.

Here are 151 Adelie penguins. Their flipper lengths average 189.95 mm, and they vary about that by 6.54 mm. The standard error of that average is 0.53 mm.

The same measurements, and the two error bars you could draw on them

Every Adelie in the dataset, then the mean with a standard deviation whisker, then the same mean with a standard error whisker. Nothing differs between the second and third except which quantity is drawn. The SD whisker spans most of the birds. The SEM whisker is smaller than the dot marking the mean.

Source Horst AM, Hill AP, Gorman KB, Zenodo, 2020. CC0 1.0. CC0 1.0

Data zenodo.3960218 · retrieved 2026-09-01 · CC0 1.0
ProcessingTwo of the 344 records have no flipper measurement and were dropped. Nothing was averaged, rescaled, trimmed or excluded as an outlier, and all 342 remaining measurements are drawn.

The SD whisker is 12.3 times the SEM whisker. Same birds, same measurements, same mean. Only the quantity on the whisker changed.

The part that should decide it for you

Take the same birds and pretend you caught fewer of them. Draw three at random, then four, then ten, and so on, and see what each statistic does.

What each statistic does as you collect more

Standard deviation and standard error against sample size, averaged over 4000 random subsamples of the same 151 birds at each size. Standard deviation is flat: the birds vary by about 6.5 mm however many you catch. Standard error falls away with the square root of n, from 3.33 mm at three birds to 0.53 mm at all of them.

Source Horst AM, Hill AP, Gorman KB, Zenodo, 2020. CC0 1.0. CC0 1.0

The standard deviation line is flat, because it is a fact about penguins and penguins did not change. The standard error line falls away, because it is a fact about your experiment.

So a figure with short error bars is not showing you a cleaner result. It may only be showing you a bigger one. Two labs could measure the same effect with the same variability and publish figures that look completely different in their apparent precision, and the difference would be how many animals they used.

That is not an argument against standard error. It is an argument for saying which one you drew, because a reader who guesses wrong will misjudge the result in a predictable direction.

There is a smaller lesson in that flat line too. At three birds the standard deviation comes out at 5.762 mm rather than 6.54: with a handful of measurements you tend to underestimate the spread as well as overstate the precision.

What each one is for

Cumming, Fidler and Vaux set this out in The Journal of Cell Biology in 2007, and it remains the reference worth reading.

Standard deviation describes your sample. Use it when the question is how variable the thing you measured actually is.
Standard error estimates how much the mean would move if you ran the experiment again. It is about the estimate, not the organism.
A confidence interval answers the question most readers think an error bar is answering, which is the range the true mean plausibly sits in. It is usually the more useful of the inferential two, and it is not much harder to produce.

Their first rule is the one nobody argues with, and it is the one most often broken: say in the legend what the error bars are. A whisker with no definition is not a measurement, it is a decoration.

What to do

Show the points where you can. With the sample sizes typical in this field, every measurement fits on the page, and then the error bar is a summary of something the reader can already see rather than a substitute for it.
Name the quantity in the legend, every time, along with n and what n counts.
Prefer a confidence interval to a standard error if you are making a claim about the mean, because it is the quantity a reader will try to read off anyway.
Never compare two figures’ error bars across papers without checking both n and both definitions.

This is on the pre-submission checklist as two lines, and A figure is an instrument covers what a bar chart does to a reader before any whisker is added to it.

Jet

Hannah Isles — Thu, 10 Sep 2026 00:00:00 GMT

The case against the rainbow is usually made by assertion. It is prettier than it is useful, it is not perceptually uniform, everyone should stop. All true, and none of it tells you what it is doing to your figure.

Here is one confocal plane of mouse embryo nuclei, deposited under CC BY in the Image Data Resource. The same array of numbers, mapped through two colourmaps, with nothing else changed.

jet

viridis

One plane, two colourmaps

Channel 1, DAPI, of image 6001240 from study idr0062. Identical data, identical display range, identical everything except the map from value to colour. Jet gives each nucleus a red core that reads as a hot spot; those cores are the same values viridis renders as a smooth continuation of the gradient around them.

Source Blin G, Sadurska D, Migueles RP, Chen N, Watson JA, Lowell S, PLoS Biology, 2019. CC BY 4.0. CC BY 4.0

Data idr0062A/6001240 · retrieved 2026-09-10 · CC BY 4.0
ProcessingOne confocal plane, z=120, channel 1 (DAPI), rescaled from the 0.5th to the 99.95th percentile of its own values and mapped through two colourmaps. Nothing else differs between the two images.

Look at the nuclei in the left-hand image. Each has a red centre with a fairly sharp edge to it, and your eye reads that edge as a boundary between two things. There is no boundary. The values slide smoothly from the rim to the middle, and you can see them doing it on the right.

Why it happens

A colourmap has one job: turn a number into a colour so that the reader can turn it back into a number. Nearly all of that work is done by lightness. Hue is decoration on top of it, which is why a sequential map has to get brighter the whole way up and never dim.

How bright each colourmap is, along its length

Relative luminance against position in the map. Viridis climbs from 0.02 to 0.78 and never falls back. Jet reaches its brightest at 65% and then drops to 0.05, which is almost exactly where it started. The bands you can see in the image above are the corners in this line.

Source sci-vis.com, 2026. Original work for sci-vis.com.

Two things follow from that shape, and both are measurable.

The steps are not even. Where the line is steep, a small change in value makes a large change in lightness, and the reader sees an edge. Where it is flat, a large change makes none, and the reader sees nothing. Summed across the whole map, jet reverses direction by 97.9% of its own lightness range. Viridis reverses by 0.07%. That is the same measure this site screens its own sequential palettes with, and anything above a few percent fails.

The top and the bottom are the same colour. Not the same hue, obviously, but lightness is what survives. Jet ends at luminance 0.05, having started at 0.02. Print it, photocopy it, or read it with no colour vision at all, and the highest values in your data are as dark as the lowest.

What that costs, in greys

Here are the same two images with the colour taken out, which is what a greyscale printer does and what a reader with achromatopsia has always had.

jet, luminance only

viridis, luminance only

The same two images, as luminance only

Converted to relative luminance, the measure that survives greyscale printing. In the jet version the nuclei have gone hollow: the red centres, which were the highest values, are now darker than the ring around them, so the structure reads inside out. Across this image jet separates the brightest and darkest one percent of the data by 54 grey levels out of 255. Viridis separates them by 145.

Source Blin G, Sadurska D, Migueles RP, Chen N, Watson JA, Lowell S, PLoS Biology, 2019. CC BY 4.0. CC BY 4.0

The nuclei have turned inside out. The brightest signal in the data is now the darkest part of each nucleus, because the highest values were red and red is dark. A reader could describe that image accurately and be wrong about the biology.

One honest complication

We first built these images with the display range set to the plane’s own minimum and maximum, which is the obvious thing to do. On that version jet came out ahead: separation of 135 grey levels against viridis on 104.

That result was an artefact. A single hot pixel set the ceiling, so almost nothing in the image reached the top of the colourmap and jet’s dark red end was never used. Switching to the percentile display range this site already uses on this study, which exists precisely so one pixel cannot set the ceiling, gives the numbers above.

Both runs are honest measurements. Only one of them measured a picture anybody would publish, and we would not have known which without checking. That is the real argument against jet, and it is not aesthetic: whether the map ruins your figure depends on where your data happen to sit inside it, and you cannot tell by looking.

What to use instead

Viridis, cividis, magma or any of the scientific colour maps published with Crameri, Shephard and Heron’s 2020 paper. They are in matplotlib, R, Fiji and Napari already, and switching is one argument to one function call.

If you want them as ready-made files, the palettes page has viridis and cividis as Fiji lookup tables, and the pre-submission checklist has this on it as one line.

What your plotting library gives you by default

Hannah Isles — Thu, 10 Sep 2026 00:00:00 GMT

You did not choose your colours. You plotted three conditions, your tool gave each one a colour, and you got on with the paper. That is the right thing to have done: the default is a decision somebody else already made, and the whole point of a default is that you should not have to think about it.

It is worth knowing what was decided.

The most ordinary figure there is

matplotlib’s shipped configuration sets its colour cycle to tab10, and tab10 begins blue, orange, green. Draw three conditions and those are the three you get. They are the same ten values as D3’s schemeCategory10, which is what this page reads them from.

In ordinary vision the orange and the green are #ff7f0e and #2ca02c, and they are 28.6 apart on the OKLab scale this site screens its own palette with. Comfortable. Under simulated protanopia they are 0.7 apart.

The second and third default colours, and the same pair to a red-blind reader

Simulation after Machado, Oliveira and Fernandes (2009) at full severity. Series two and three of matplotlib's default cycle separate by 28.6 in ordinary vision and by 0.7 under protanopia, against the 9.5 floor this site holds its own six colours to. They are not close. They are the same colour.

Source sci-vis.com, 2026. Original work for sci-vis.com.

What that does to a chart

Here are 342 penguins, three species, drawn the way a default draws them: one axis, colour carrying the grouping and nothing else carrying it. On the left as most readers see it. On the right as a protanopic reader does.

Reproduction. Reproduces matplotlib's first three default colours so the piece can show what they do, rather than redrawing them in a palette that would hide the problem. It is drawn to the original's rules, not ours, and our own checks refuse one thing in it.

As drawn

Reproduction. The same three default colours after protanopia simulation, which is the point of the figure and cannot be shown in this site's own palette. It is drawn to the original's rules, not ours, and our own checks refuse one thing in it.

To a protanopic reader

Three species, distinguished by colour alone

The same measurements twice, in matplotlib's first three default colours and in those same colours under simulated protanopia. On the right there are two colours, not three: the orange group and the green group have merged, and no amount of squinting separates them because there is nothing else to separate them by.

Source Horst AM, Hill AP, Gorman KB, Zenodo, 2020. CC0 1.0. CC0 1.0

Roughly one man in twelve has some form of red-green deficiency, and protanopia is one of the two common kinds. On a paper with four reviewers and a few hundred readers, somebody in that chart’s audience is seeing two groups.

It is not only matplotlib, and it is not only other people

The useful question is not whether a palette is perfect but how far you can get before it stops working. Series are handed out in order, so what matters is how many you can draw before two of them collide.

How many series each default gets through

Series drawn before two of them fall below OKLab 9.5 under normal vision, protanopia, deuteranopia or tritanopia. Two of these tools survive four series. matplotlib does not survive three. The bar for this site is six, which is where it stops on principle rather than where it fails.

Source sci-vis.com, 2026. Original work for sci-vis.com.

Observable Plot is on that chart because it draws every figure on this site. Its default gets through four series, which is better than matplotlib and is still not a palette anybody checked for this. We do not use it: everything here comes from src/lib/palettes.ts, which was searched against exactly this test, and the build refuses any colour that did not come from there.

There is a second problem in the same data that has nothing to do with colour vision. matplotlib’s orange is 2.45:1 against a white page. A mark needs 3:1 to be legible as a thin line, so that colour is already too faint for an error bar or a hairline before anybody’s colour vision is considered. Tableau’s default has three such colours in its first six. Observable Plot’s has five.

What to do

Nothing here requires you to design a palette. It requires you to stop accepting one.

Set the colours explicitly, once, in whatever your lab plots with. Okabe and Ito’s set is the standard answer, it is free, and it is one line of configuration.
Never let colour be the only difference between two series. Shape, dash or a direct label costs nothing and makes the question moot.
If you must keep a default, keep it to two series. Every palette here survives two.

The pre-submission figure checklist has this on it, and the palettes page has ready-made hex lists and Fiji lookup tables if you would rather take one than pick one.

What your journal checks, and what it does not

Hannah Isles — Mon, 07 Sep 2026 00:00:00 GMT

You are a few days from submitting. You open the author guidelines, find the figure section, and it tells you 300 dpi, TIFF, and a width in millimetres. You make your figures 300 dpi, TIFF, and that width. You have now met every requirement your journal stated, and you have no idea whether your figures work.

That is not a criticism of you. It is a description of the lists. They are production specifications, written so the typesetter’s software does not choke, and they have been mistaken for a standard of quality because they are the only written thing anybody hands you.

Here is what two large open-access publishers ask for. Both were read from the publishers’ own pages on 7 September 2026, and both are quoted rather than summarised, because the pages that rank for these searches are mostly written by companies selling figure software and their numbers are not reliable.

The same resolution, a different idea of a figure

PLOS ONE wants 789 to 2250 pixels wide at 300 dpi, which it also gives as 6.68 to 19.05 cm. TIFF or EPS, nothing else. 300 to 600 dpi. Text in Arial, Times or Symbol only, at 8 to 12 point. Under 10 MB, LZW compressed, flattened, no alpha channel.

Frontiers wants 300 dpi at final size, one column at 85 mm or two at 180 mm, TIFF or JPEG or EPS. The smallest visible text no less than eight points when viewed at actual size. No line thinner than two points. It warns you off red and green as your two indicators, notes that more than 99% of colour-blind readers have a red-green deficiency, and asks you to carry the distinction in shape, labels or size as well as colour. It then gives contrast ratios: 4.5:1 for level AA, 7:1 for AAA.

Read those twice. Both demand 300 dpi. One of them tells you which three fonts you may use and says nothing whatsoever about colour. The other tells you the contrast ratio your figure has to reach.

The requirements you are handed are not a standard. They are a local dialect, and the parts that decide whether your figure can actually be read are the parts most journals leave out.

Three things that go wrong, at the size they go wrong

Every specification below is stated at final size, and that phrase is doing almost all of the work. A figure is not published at the size you drew it.

Your type is not the size you set it

Frontiers asks for eight points minimum at actual size. If you build a figure at the two-column width and it runs at one column, everything in it is multiplied by 85/180, which is 0.47.

Label type, at the size it will be printed

Set at true physical size. The eight point minimum is the line to clear after reduction, not before it. A twelve point label drawn for the two-column width lands below the minimum when the figure runs at one column.

Source sci-vis.com, 2026. Original work for sci-vis.com.

The fix costs nothing and almost nobody does it: decide the column width before you draw, and set the type in the figure so it is right at that width. Drawing large and letting the journal shrink it is the single most common way a perfectly good figure arrives illegible.

Your lines are thinner than the minimum

Frontiers will not take a line under two points. That is thicker than it sounds, and it is very likely thicker than what your tool gave you. Observable Plot, which draws every chart on this site, defaults to a stroke of one pixel, which is 0.75 point: we checked our own tool rather than assume, and it comes in at a little over a third of the minimum. Check yours the same way, in whatever units it reports, before you take a default on trust.

Stroke weights, at the size they will be printed

At true physical size. The default weight in most charting tools is around three quarters of a point, which is roughly a third of the minimum Frontiers will accept, and the thinnest rules here are the ones that disappear when a figure is printed rather than viewed.

Source sci-vis.com, 2026. Original work for sci-vis.com.

Your two colours may be one colour

This is the one that no production specification catches, and it is the one that costs a reader the result.

Take a mid red and a mid green, the pair any tool will hand you for two conditions. In ordinary vision they are far apart: 33.8 on the same OKLab scale this site uses to screen its own palettes, where anything below about ten is where two series stop being reliably distinguishable. Under simulated deuteranopia, the most common form of colour blindness, the same two colours fall to 4.1.

A red and a green, and the same pair as a deuteranopic reader sees them

Simulation after Machado, Oliveira and Fernandes (2009) at full severity, using the same code that screens this site's own palettes. The pair separates by 33.8 in ordinary vision and by 4.1 under deuteranopia, against a floor of 9.5 that this site holds its own six colours to. A figure using these two colours meets every specification PLOS ONE states.

Source sci-vis.com, 2026. Original work for sci-vis.com.

Nothing in the PLOS ONE figure specification would stop that figure. It can be a 300 dpi TIFF, 19 cm wide, set in 9 point Arial, under 10 MB, and every stated requirement is met. Roughly one man in twelve cannot read it.

What to do instead

The production specification is a floor, and you should meet it. Then check the four things it does not ask about.

Decide the column width first, and set your type so it clears eight points at that width, not before reduction.
Thicken your strokes. Whatever your tool gave you, it is probably a third of what a printer wants.
Never let colour be the only difference between two series. Shape, dash pattern or a direct label costs you nothing and survives both colour blindness and a photocopier.
Export vector where the figure is drawn rather than photographed. A plot is lines and text, and dpi is meaningless for it. Rasterising a chart to hit a dpi number throws away the thing that made it sharp.

None of that is in most author guidelines. All of it is what an editor notices.

The four checks above, and the twenty-two that go with them, are on one page at the pre-submission figure checklist, which you can print and keep beside you while you build the figures. It is free and there is no email wall.

A figure is an instrument

Hannah Isles — Sat, 05 Sep 2026 00:00:00 GMT

There is a habit of speaking about figures as though the choices in them were matters of taste. A reviewer prefers the bar chart. An editor finds the colours busy. A supervisor has always used pie charts and sees no reason to stop.

The habit is comfortable and it is wrong, and it has been demonstrably wrong for about forty years. Which encoding you choose changes how accurately people read the numbers, by amounts that have been measured in controlled experiments and replicated since. A figure is not decoration applied to a result. It is the instrument through which the reader takes the measurement, and instruments can be calibrated.

The reader’s instrument, calibrated

Cleveland and McGill put this on a formal footing in 1984. They set out a list of elementary perceptual tasks, the small judgements a reader makes when decoding a chart, and ran experiments to find how accurately people perform each one. Position along a common scale came out at the top. Angle and area came a long way below it.

That is a statement about people, not about charts. It is why a dot plot and a pie chart of the same numbers are not two presentations of one result.

An interactive self-test. It shows two quantities encoded as positions on a scale, as lengths of unaligned bars, as angles of pie slices, or as areas of circles, and asks the reader to estimate the smaller as a percentage of the larger. Without JavaScript it shows the published ranking instead: position on a common scale is read most accurately, then length, then angle, then area.

What the studies found

Cleveland and McGill ranked elementary perceptual tasks by how accurately people read them. Position on a common scale came top. Angle and area came far below it. That ordering, rather than any particular number, is the finding:

Position on a common scaletwo dots on one axis
Length, not alignedtwo bars with different baselines
Angletwo slices of a circle
Areatwo circles

There is a self-test here that lets you try the four encodings yourself. It needs JavaScript, which is not running. The findings above are the point of it and do not.

Four ways of showing the same comparison

The same judgement, asked four ways. Position on a common scale is the one people read most accurately; angle and area are much worse. Your own twelve attempts prove nothing, which the results say plainly.

Source sci-vis.com, 2026. Original work for sci-vis.com.

What a bar does to a reader

The most-used chart in the life sciences is a bar of a group mean with an error bar on top, and the case against it is usually made on the grounds that it hides the distribution. That is true and it is the smaller half of the problem.

Newman and Scholl reported in 2012 that readers do something stranger. Shown a bar representing an average, people judge a value that falls inside the bar as more likely to be a real observation than a value the same distance from the mean but outside it. The bar is a shape with an inside and an outside, and readers treat that boundary as if it meant something. It does not. The distribution is symmetric about the mean and the bar’s edge is an artefact of the encoding.

So a bar chart does not merely fail to show the spread. It puts a boundary in the reader’s mind where the data have none.

A bar chart of three species means wipes to reveal every individual measurement on the identical axis. All 342 birds sit between 172 and 231 millimetres, in the top fifth of the plot, while the bars fill the whole height from zero.

The same axis, with the bars replaced by the measurements

Every bird sits in the top fifth of the space the bars were filling. The bars are not wrong. They are almost all of the ink for almost none of the information, and the shape they make is read as a boundary.

Source Horst AM, Hill AP, Gorman KB, Zenodo, 2020. CC0 1.0. CC0 1.0

The error bar is not helping either

The usual defence of the bar chart is that the error bar restores what the bar left out. Correll and Gleicher tested that in 2014 and found the error bar carries its own misreadings, and that other encodings of the same uncertainty were read more accurately.

There is a plainer problem underneath. A standard error is a statement about how precisely the mean was estimated, and it shrinks as the sample grows. A standard deviation is a statement about how much the things themselves differ, and it does not. Those answer different questions, and only one of them is usually on the chart.

The bar chart, with the parts that are doing no work

About nine tenths of each bar sits below the shortest flipper measured in that species, where there are no data at all, and the error bar is a quantity most readers will misread.

The outlined area is the part of the bar below the shortest flipper measured in this species, 172 mm. It is about nine tenths of the bar, and not one of the 151 birds is in it. This is the region Newman and Scholl found readers treat as more probable than the space just outside the bar.
The error bar is here, and it is about half a millimetre. Standard error shrinks as the sample grows, so on 151 birds it says almost nothing about how much the birds differ from each other.
The axis starts at zero, which is correct for bars, and the consequence is that the whole interesting range is squeezed into the top fifth of the plot. Bars need a zero baseline, and a zero baseline spends the plot on a quantity that never goes near zero.

Source Horst AM, Hill AP, Gorman KB, Zenodo, 2020. CC0 1.0. CC0 1.0

Drawn as position instead

Once the bars are gone the axis no longer has to start at zero, the plot can spend its height on the range the data occupy, and every bird is placed rather than summarised. This is the same comparison, encoded as position on a common scale, which is the encoding Cleveland and McGill found people read most accurately.

Flipper length by species, every bird drawn

Adelie and Chinstrap overlap across most of their range. Gentoo barely touches either. That is a different sentence from the means differ, and it is usually the one a reader wants.

Source Horst AM, Hill AP, Gorman KB, Zenodo, 2020. CC0 1.0. CC0 1.0

Why this is a science and not a style

Franconeri, Padilla, Shah, Zacks and Hullman reviewed this literature in 2021 for Psychological Science in the Public Interest, and the title is the argument: the science of visual data communication, what works. There is a body of experimental results here, with methods, effect sizes and replications, and it is not a collection of opinions about what looks nice.

Worth noting, on a site that publishes its own corrections prominently: that review issued a corrigendum the following year. A field that corrects itself in public is behaving like a science, which is rather the point.

None of this means a figure cannot also be beautiful, or that judgement stops mattering. It means the judgement has something to answer to. When someone says they prefer the bar chart, there is a reasonable next question, and it is not about preference.

What changed, and what did not

Bars replaced with one point per bird. Nothing is aggregated away.
Points positioned by a deterministic dodge rather than random jitter, so the figure is identical on every build and two readers see the same picture.
Standard error bars removed. The mean is drawn as a rule across each group and labelled, and the spread is visible directly rather than summarised.
The y axis no longer starts at zero, which is legitimate once a mark’s position rather than its length carries the value.
Species separated into panels, which makes room for the points to spread.
One colour for every point. The panel and its label already identify the species, so hue was restating the axis.
Group sizes written on the figure. Three groups of 151, 68 and 123 birds should not look equally certain, and on the bar chart they did.

What did not change is the data. Every figure here is drawn from the same 342 measurements, and the point of the piece is that this is exactly the problem: the numbers were never in question.

References

Cleveland WS, McGill R. Graphical perception: theory, experimentation, and application to the development of graphical methods. Journal of the American Statistical Association 1984. 10.1080/01621459.1984.10478080
Newman GE, Scholl BJ. Bar graphs depicting averages are perceptually misinterpreted: the within-the-bar bias. Psychonomic Bulletin & Review 2012. 10.3758/s13423-012-0247-5
Correll M, Gleicher M. Error bars considered harmful: exploring alternate encodings for mean and error. IEEE Transactions on Visualization and Computer Graphics 2014. 10.1109/tvcg.2014.2346298
Heer J, Bostock M. Crowdsourcing graphical perception: using Mechanical Turk to assess visualization design. CHI 2010. 10.1145/1753326.1753357
Franconeri SL, Padilla LM, Shah P, Zacks JM, Hullman J. The science of visual data communication: what works. Psychological Science in the Public Interest 2021. 10.1177/15291006211051956. Corrigendum 2022, 10.1177/15291006221099494
Horst AM, Hill AP, Gorman KB. palmerpenguins: Palmer Archipelago (Antarctica) penguin data. 2020. 10.5281/zenodo.3960218. CC0