They are probably coming out wrong because your dpi changes. '.' attempts to fill in a pixel, but pixels are sneaky and change their size. Better to not use '.' if you care about exact sizes.

They are probably coming out wrong because your dpi changes. '.' attempts to fill in a pixel, but pixels are sneaky and change their size. Better to not use '.' if you care about exact sizes.

I'm not actually sure what "." actually attempts to draw, since the pixel marker according to matplotlib.markers should be a comma: ",".

But even if it's supposed to fill in a certain number of pixels, shouldn't those numbers of pixels be the same in the legend and the plot?

I'm not sure this is specific to .; the o result looks slightly off as well.

I think that 'o' is an optical illusion - it looks fine when you line them up.

The problem here is the peculiarity of how '.' is created - it is really just a circle with the size reduced by 0.5. I don't understand why this was done, or why this marker was even created. It doesn't seem worth the bother of tracking this reduction into the Line2D object. One could rewrite '.' to do the right thing without relying on circle, but - I would tend to say that if you want the legend to work, just use 'o' since it is exactly the same as '.'

Should we then disallow '.' as scatter marker?

Perhaps some relevant context here on how the concept of "marker size" across different marker shapes is rather complicated: #15703

Sure, but those markers should be consistent in the legend. The problem here is that the size gets halved for '.' but it is still a circle, so when the legend gets made, the marker is sized like it is 'o'.

OK, but the half size reduction for '.' is set by reducing the transform on MarkerStyle._transform from the original circle one. But this transform is used for all the triangle directions (as long as they're not half-filled) and that seems to be okay?

For reference, here are all filled styles:

Not all of them match exactly, but that's most likely from snapping and antialiasing (importing into GIMP and thresholding exactly white pixels shows that some markers differ by 1 pixel or so, while many are exact same size.) But from what I can tell, both 100 diamonds are not the same size (17x17 in legend vs 22x22 in axes).

And for unfilled styles:

In addition to the 'o', all the numbered styles are very wrong in the legend.

import matplotlib as mpl
import matplotlib.pyplot as plt
from matplotlib.markers import MarkerStyle


print(mpl.__version__)


def do_plot(ax, marker, fillstyle):
    points = ax.scatter([0, 0, 0], [1, 0, -1], s=[10, 50, 100],
                        marker=MarkerStyle(marker, fillstyle=fillstyle))

    # add legend
    handles, labels = points.legend_elements(prop="sizes", num=None,
                                             color='C0')
    ax.legend(handles, labels, loc='right')
    ax.text(-0.5, 0, f'{marker!r} {fillstyle}', va='center', ha='center')

    ax.set(xticks=[], xlim=(-1, 1), yticks=[], ylim=(-2, 2))


# Filled markers
fig, ax = plt.subplots(len(MarkerStyle.filled_markers),
                       len(MarkerStyle.fillstyles),
                       figsize=(12, 10.8))
for i, marker in enumerate(MarkerStyle.filled_markers):
    for j, fillstyle in enumerate(MarkerStyle.fillstyles):
        do_plot(ax[i, j], marker, fillstyle)
fig.supylabel('Filled markers', ha='right')
fig.subplots_adjust(left=0.02, right=1, bottom=0, top=1, wspace=0, hspace=0)

# Unfilled markers
unfilled_markers = sorted(
    (k for k, v in MarkerStyle.markers.items()
     if k not in MarkerStyle.filled_markers and v != 'nothing'),
    key=str)

unfilled_count = len(unfilled_markers)
nrows = round(unfilled_count / len(MarkerStyle.fillstyles) + 0.5)
figheight = 10.8 / len(MarkerStyle.filled_markers) * nrows
fig, ax = plt.subplots(nrows, len(MarkerStyle.fillstyles),
                       figsize=(12, figheight))
for i, marker in enumerate(unfilled_markers):
    do_plot(ax.flat[i], marker, 'full')
for a in ax.flat[unfilled_count:]:
    a.remove()
fig.supylabel('Unfilled markers', ha='right')
fig.subplots_adjust(left=0.02, right=1, bottom=0, top=1, wspace=0, hspace=0)

plt.show()