# Which image format is best to use in a scientific paper?

When submitting a scientific paper written using LaTeX, to a journal, which high-quality image format do you usually use?

• I always try to use PDF. But one time the journal, even using Latex, ask for TIFF file instead.
– Camps
Sep 10 at 12:45
• Meta discussion relevant to this question: mattermodeling.meta.stackexchange.com/q/291/671 Sep 11 at 16:27
• @Vikki I believe they mean what image format is used for the embedded images
– Tyberius
Sep 12 at 2:43
• @ChiKou Hm. If you submit a single PDF, isn't it your choice how you embed your images? Or do you submit pure text as a PDF with gaps for images submitted separately? The arXiv submission mentioned in the accepted answer has both a rasterized image on p. 9 and the vector graphic shown in that answer on p. 10. I suppose it's a matter of the image source and the Latex toolchain what the resulting PDF contains, and not up to the editor (except insofar as the editor may decide to (badly) rasterize the entire publication). Sep 12 at 10:57
• Also, in order to print digitally on paper the publication must be rasterized in any case; the question is probably how well that is done (depending on the printing hardware and, again, the toolchain). Sep 12 at 10:59

For scientific papers in matter modeling, I always use vector graphics (e.g. .eps, .pdf, or .svg, though the journals will not typically accept .svg format).

I stopped publishing in ACS journals (e.g. JCTC and JACS) after they turned one of my vector graphics into a pixelated raster, before publishing it. Even though it was 2015, and I'd been publishing papers with vector graphics in journals like JCP for many years already, they said they don't accept EPS or PDF files, when I was publishing this paper in JCTC. In Figure 2, you can see the difference between the published version (top) zoomed in as much as possible (and yet still pixelated) and the version on arXiv (bottom) which can be zoomed in much more, and still doesn't get pixelated (click the image to see it larger):

As Camps explained in his comment, they only accepted TIFF files at the time:

"Date: Tue, 9 Jun 2015 at 10:01
Subject: RE: FW: ct-2014-01066k Why Quantum Coherence Is Not Important in the FennaMatthewsOlsen Complex

Dr. Dattani,

Thank you for your message concerning this manuscript. We are currently exploring the use of vector graphics, but at this time we are unable to grant your request. If you wish to see if tif files submitted as explained below would improve the resolution of your figures, you may send new files directly to me by e-mail. Please advise.

Thank you,

Production Assistant
American Chemical Society"

That was in response to an email in which I'd complained that the label for picoseconds in my original vector graphic (bottom of image below) had been pixelated by the journal when they converted it to a raster (top of image below), here is a screenshot of that email I sent:

Buttonwood made some very good points this comment, with which I completely agree. Vector images have searchable text whereas raster images typically do not. Furthermore the size of the image will be much smaller if you use a vector graphic instead of a raster: arXiv has a limit on the size of the files submitted, and I once had a submission that was at least 10x too big in file size to get through arXiv's submission portal, but when I converted the figures to PDF format, the entire paper's size was about 53kB.

• Not only crispiness is a difference. Disregarding illustrations and the searchable text layer, possibly file volume of the .pdf using bitmap fonts may be larger than one using postscript fonts. Sep 10 at 13:30
• Although it is of course true that vector formats are preferrable for, well vector graphics, it is not true that they are better for all figures. Heatmaps generally require some kind of rasterization regardless. Tools that export to .eps typically do that under the hood, but often less than optimally. .png with a suitably optimised palette can be much more efficient, with negligible visual difference even in print. Sep 11 at 17:54
• @leftaroundabout I think you should expand that into an answer, that presents .png as an alternative to my answer. I'd actually agree with it for some cases! Sep 11 at 19:26
• @leftaroundabout No working code with me for this one, but a colleague of mine was plotting heatmaps with gnuplot, exported as crisp .eps files. Since he knew 1) they would be print anyway on white paper, 2) only two thirds only would be non-zero intensity in the projection requiring a «tile» at all to plot the map, he used a conditional Boolean to leave many 0-grid pixels just undefined. Equally, only the top most surface of the heat map was plot instead an averaging layer approach previously used. Thus, with .eps only about 25% to 50% of the .png, the .pdf of the SI was small enough for ACS. Sep 13 at 16:51
• It would be worth the archeology to check if, e.g., matplotlib (or your plotter of choice) equally knows such a «conditional tiling» to plot pixels only if they exceed a certain threshold. Sep 13 at 16:53

Complementary to the answer given by Nike Dattani with focus on file format, I would like to draw your attention to how you present the content visually in photos, schemes and diagrams known as artwork.

• Regardless of the file format eventually elected (where suitable, vector based like .(e)ps or .pdf), determine in advance the dimension of the illustration: Are you invited to design the outer/inner title of the issue of the journal, may you use color(s), will the illustration span over two columns, or are you constrained to the width of a single narrow column. This may be a guide to balance the content and detail it contains.

The blueprints by architects, datavisualizations in the newspapers, or work e.g., by Edward Tufte may be an inspiration to remove clutter and reorganization e.g., if the illustration is a photography. Beside inspiration by illustrations by your colleagues, well maintained scientific journals provide advice, too (example IUCr).

• If the illustration isn't a photography, design it like a flag -- easy recognizable from distance (e.g., Chicago):

1. Keep It Simple.
2. Use Meaningful Symbolism.
3. Use 2 or 3 Basic Colors.
4. (...)
5. Be Distinctive or Be Related.

(Credit to Good Flag, Bad Flag compiled by NAVA. I do not think their point 4 is relevant here.)

So if you plot data in a diagram, consider markers clearly distinctive by color, yet still distinctive by shape if the publication isn't transmitted in color passing a Xerox:

Note: If you start with a too large dimensions, the journal's request to scale the illustration to fit into a single column may render these markers difficult/impossible to discern (especially if a publisher converts them into pixel bitmaps).

• The more discrete colors in the illustration, the more the potential lost of information when this is reproduced in grey scale:

Use freely accessible tools like ColorBrewer2 to probe color schemes safe enough to convey sequential/diverging, or qualitative data. You may toggle on/off options for greyscale, photocopy, color blindness.

• Consider color blindness. There are much more men (and to lesser extent, women) prone to this, than you might think (see examples). Check your illustration in advance with tools like visicheck, or your image editor (example GIMP).

• If your illustration is a 2D projection of 3D map where e.g., intensity of a property is color encoded, substitute default color schemes like «jet» by one like «viridis» (continous data):

(credit)

or «blue-white-red» about divergent data by «bent-cool-warm»:

(credit)

It is worth to invest some time and effort into this topic. The work by Kenneth Moreland (a presentation), or Kristen Thyng may guide you further. Often, such palettes are freely available for e.g., Python's mathplotlib, or gnuplot; so, once set up well, it will work in the background for you.

It isn't easy to follow all these points at once equally well. Like any profession, it asks for training, curiosity, and testing.