When an editor opens a new submission they read the abstract and they look at the figures. Often in that order, often within a minute or two, and usually before reading a word of the introduction.
This is not laziness. Figures are the most information-dense part of a paper and the least rhetorical. Prose can describe an effect as striking; a panel either shows it or does not. An experienced reader goes to the figures because that is the fastest honest route to what the paper has.
Which means the figures are carrying far more of the first impression than most authors assume, and they are usually the last thing finished.
What goes wrong most often
Too much in one panel. A figure that contains four comparisons is making the reader decide which one matters. Deciding that is your job. The most common single improvement to a figure is removing something from it.
No stated claim. Every panel should be answering a question, and the caption should say what the answer is rather than only what was plotted. A caption that reads as a methods description leaves the reader to extract the finding themselves, and some of them will extract a different one.
Encoding that fights the comparison. If the point is a difference between two groups, the figure should put those two things next to each other on a common scale. Grouped bars across four panels with different y-axes make the comparison you care about the hardest one in the figure to make.
Bars where the distribution is the point. A bar with an error cap hides the shape of the data and the sample size. If n is small, plot the points. Readers have become alert to this and an editor noticing that a distribution has been hidden is a bad moment, because the obvious question is what it was hiding.
Error bars that are not identified. Standard deviation, standard error and a confidence interval are three different claims and they look identical on the page. If the caption does not say which, the figure cannot be assessed, and a reviewer will ask.
Significance markers without effect sizes. Asterisks tell the reader a comparison cleared a threshold. They say nothing about how big the effect is, which is usually the thing that matters.
Colour
Colour is where good scientific figures most often fail, and the failures are systematic rather than matters of taste.
Rainbow and jet colour maps are still common and still misleading: they are not perceptually uniform, so they invent boundaries where the data are smooth and flatten real differences elsewhere. For continuous data use a sequential map designed to be uniform, and a diverging one only when there is a genuine midpoint.
Red and green together will be unreadable for a meaningful fraction of your readers. In microscopy this is an unforced error, because the channels can be shown in magenta and green, or as separate greyscale panels, and lose nothing.
And colour should not be the only thing carrying a distinction. If the only difference between two lines is their hue, the figure fails in print, in greyscale, on a projector, and for colour-blind readers, which between them is most of the ways it will be looked at.
Microscopy specifically
Scale bars on every image panel, with the size stated in the caption. Any adjustment to brightness or contrast declared, applied to the whole image, and applied identically across conditions, which is the part people get wrong under time pressure. If an image is described as representative, be able to say what it is representative of, and show the quantification somewhere. A single beautiful field is an illustration, not evidence, and editors read it as one.
Graphical abstracts
A graphical abstract is a single image summarising the paper, used in search results, tables of contents and social posts. Journals increasingly ask for one and many authors produce it in an hour on the day of submission.
It is worth more thought than that, because it is often the only part of your paper most people ever see.
The common failure is treating it as a figure. It is not: a figure supports a claim for a reader who is going to read the paper, and a graphical abstract makes one claim to a reader who is probably not. So it should contain one idea, legible at the size of a thumbnail, with no panel that requires the caption to make sense. Text should be minimal and readable, which in practice means a handful of words at a font size that survives being shrunk.
The second failure is a diagram of the experiment rather than the finding. A reader does not need your workflow. They need to know what you found.
Check the journal’s requirements before building it, because dimensions, file format and whether text is allowed vary, and a rebuild at submission is an avoidable annoyance.
Where to go deeper on this
I also edit sci-vis.com, a publication about visualising science that takes published figures and either annotates what is going wrong in them or replots them from the deposited data to show what the figure could have said. It is a slower, more detailed treatment of everything on this page, with the arguments made against real examples rather than in the abstract. The piece on rainbow colour maps is a reasonable place to start, and there is a pre-submission figure checklist written to be kept.
The underlying rule there is worth stating here too, because it is the line that separates presentation from misconduct: you may change how a figure encodes, labels, aggregates and frames the data. You may not change what the data say. If redrawing a figure makes the effect look larger or cleaner than it is, the redraw is wrong, however much better it looks.