With the new DejaVu font default many of the symbols (most notably the integral, product, and summation symbol) are very small, as they are designed for inline math. The DejaVu fonts include display-sized variants for some of the math symbols outside of the unicode range. According to @mdboom, it is currently non-trivial to access these variants in a consistent manner from matplotlib, so here I did a bit of hack so we can always access them:

• I modified the subset.py script from google web fonts so that it can create a font with a subset of all the glyphs with names ending in .display and move them to the unicode index corresponding to the regular glyph. I created two ttf files with this: DejaVuSansDisplay.ttf and DejaVuSerifDisplay.ttf. The family names have an additional Display at the end to differentiate them from the regular fonts. These are pretty small, as they only have a dozen characters.
• I have saved this modified subset.py with the fonts in lib/matplotlib/mpl-data/fonts/ttf. I think it's good to keep it so this modification can be done in the future if we update the version of DejaVu fonts, but I am not sure this is the right place. A better location might be tools. Ideas? It has an Apache license, so AFAICT it's ok to distribute it.
• The DejaVu font classes in mathtext.py will now look first for a glyph in the display version of the fonts (I have added them as fontname ex, following the cmex example), and if not found will use the regular version.

The baseline images definitely look better, specially the ones that include fractions and integrals.