The standard Arabic rendering recipe corrupts Arabic on any modern renderer - now filed upstream
arabic_reshaper + python-bidi is the advice in almost every tutorial for drawing Arabic.
On a renderer that already does complex-text layout - Pillow linked against Raqm/HarfBuzz, which
is the default - it now corrupts the text, because shaping and the bidi reorder happen twice.
"welcome":
| rendered | |
|---|---|
| no preprocessing | مرحبا بكم |
| reshape() + get_display() | مكب ابحرم |
5 strings x 3 fonts per path:
| path | no preprocessing | with reshape + bidi |
|---|---|---|
| Pillow with Raqm | 15/15 identical | 0/15 - 14 broken, 1 recognisable |
| Pillow without Raqm | 0/15 - 15 broken | 0/15 - 9 recognisable, 6 still broken |
Two things follow. With Raqm, the preprocessing takes a path that is 100% correct and breaks all
of it. Without Raqm it is a partial rescue, not a fix - 9 of 15 move from broken to merely
recognisable and 6 stay broken. A pass/fail score reports that second row as "works", which is
how the recipe keeps its reputation.
Detection is one line:
from PIL import features
needs_reshaping = not features.check("raqm")
Now filed upstream with the maintainer:
https://github.com/mpcabd/python-arabic-reshaper/issues/102
Scoring is shape-invariant IoU against a reference, normalised to the ink bounding box, graded in
three bands rather than pass/fail - the middle band is what surfaced the partial-rescue case.
All renders and per-item results are in this repo, CC BY 4.0.