arshape / render_bench.py
syamjithnk's picture
Upload render_bench.py with huggingface_hub
a81d94d verified
Raw
History Blame Contribute Delete
7.01 kB
#!/usr/bin/env python3
"""ArShape — when does the standard Arabic rendering recipe break Arabic?
Author: Syamjith NK
REBUILT after the first version measured the wrong thing. Recording why, because the
mistake is the finding.
I assumed Arabic needs `arabic_reshaper` + `python-bidi` applied before drawing, and
that plain rendering would come out disconnected. That is the advice you find
everywhere. Rendering it proved the opposite: on a modern text stack, plain output was
correct (مرحبا بكم) and the reshaped output was mangled (مكب ابحرم).
The reason is that Pillow 12 links **Raqm/HarfBuzz**, which already performs shaping
and the bidi algorithm. Pre-reshaping hands it text that has been shaped once and
reordered once, and it does both again. Two rights make a wrong.
So the real question is not "does Arabic need reshaping" but:
Does this rendering path ALREADY do complex-text layout?
yes -> reshaping breaks it
no -> reshaping is required
This benchmark measures that 2x2 directly, plus the font-coverage trap that bit the
first run: SFArabic.ttf has no Latin digit glyphs, so "في عام 2026" renders the digits
as .notdef boxes while looking perfectly fine to anyone not reading the numbers.
"""
import json
from pathlib import Path
import numpy as np
from PIL import Image, ImageDraw, ImageFont
HERE = Path(__file__).parent
OUT = HERE / "out"; OUT.mkdir(parents=True, exist_ok=True)
SIZE = (900, 170)
PT = 60
FONTS = {
"SFArabic": "/System/Library/Fonts/SFArabic.ttf",
"GeezaPro": "/System/Library/Fonts/GeezaPro.ttc",
"IBMPlexSansArabic": str(Path.home() / "Library/Fonts/IBMPlexSansArabic-Regular.ttf"),
}
CASES = [
("greeting", "مرحبا بكم"),
("sentence", "التزامنا بالنزاهة"),
("with_digits", "في عام 2026"),
("mixed_latin", "شركة Pixelogik للإنتاج"),
("diacritics", "مَرْحَبًا بِكُمْ"),
]
def reshape(text: str) -> str:
import arabic_reshaper
from bidi.algorithm import get_display
return get_display(arabic_reshaper.reshape(text))
def draw(text: str, font_path: str, engine, pre_reshape: bool) -> Image.Image:
if pre_reshape:
text = reshape(text)
im = Image.new("L", SIZE, 255)
f = ImageFont.truetype(font_path, PT, layout_engine=engine)
ImageDraw.Draw(im).text((SIZE[0] - 30, 45), text, font=f, fill=0, anchor="ra")
return im
def ink_iou(a: Image.Image, b: Image.Image) -> float:
"""Shape similarity, INVARIANT to where the text sits.
A raw pixel IoU compares position as well as shape, and the two layout engines
place glyphs a few pixels apart. That made a visually IDENTICAL rendering score
zero - the metric could not tell "same words, different kerning" from "different
words". Crop each to its ink bounding box and normalise before comparing, so this
measures the letterforms and not the layout offset.
"""
def norm(im):
arr = (np.asarray(im) < 128)
if not arr.any():
return np.zeros((64, 256), dtype=bool)
ys, xs = np.where(arr)
crop = Image.fromarray((~arr[ys.min():ys.max() + 1, xs.min():xs.max() + 1]).astype(np.uint8) * 255)
return np.asarray(crop.resize((256, 64), Image.LANCZOS)) < 128
x, y = norm(a), norm(b)
u = (x | y).sum()
return float((x & y).sum() / u) if u else 1.0
def missing_glyphs(text: str, font_path: str) -> list[str]:
"""Characters the font has no glyph for. These render as .notdef boxes and are
invisible to anyone skim-reading the language they DO know."""
try:
from fontTools.ttLib import TTFont, TTCollection
f = (TTCollection(font_path).fonts[0] if font_path.endswith(".ttc")
else TTFont(font_path, fontNumber=0))
cmap = set()
for t in f["cmap"].tables:
cmap |= set(t.cmap.keys())
return [c for c in dict.fromkeys(text) if c.strip() and ord(c) not in cmap]
except Exception:
return []
RAQM = ImageFont.Layout.RAQM
BASIC = ImageFont.Layout.BASIC
PATHS = {
"raqm_plain": (RAQM, False), # modern stack, text passed through as-is
"raqm_reshaped": (RAQM, True), # modern stack + the legacy recipe
"basic_plain": (BASIC, False), # no complex-text layout, no recipe
"basic_reshaped": (BASIC, True), # no complex-text layout + the recipe
}
if __name__ == "__main__":
rows = []
for fname, fpath in FONTS.items():
if not Path(fpath).exists():
print(f" (skipping {fname}: not installed)"); continue
for cid, text in CASES:
# Ground truth: raqm_plain, verified BY EYE in the first run. Everything
# is measured against it, and the contact sheet lets a human re-check.
ref = draw(text, fpath, RAQM, False)
miss = missing_glyphs(text, fpath)
for pname, (engine, pre) in PATHS.items():
im = draw(text, fpath, engine, pre)
im.save(OUT / f"{fname}__{cid}__{pname}.png")
iou = ink_iou(im, ref)
rows.append({"font": fname, "case": cid, "text": text, "path": pname,
"iou_vs_reference": round(iou, 3),
# Three bands, not a pass/fail. Antialiasing across two
# layout engines means even a visually identical render
# will not reach 0.97, and a single threshold hid the
# most interesting result: the legacy recipe RECOVERS
# plain Arabic but still fails on digits, Latin runs and
# diacritics, because those need real shaping regardless.
"verdict": ("identical" if iou > 0.95 else
"recognisable" if iou > 0.60 else "broken"),
"missing_glyphs": miss})
(HERE / "results.jsonl").write_text(
"\n".join(json.dumps(r, ensure_ascii=False) for r in rows) + "\n")
print(f"\n{'path':17}{'identical':>11}{'recognisable':>14}{'broken':>9} per-case IoU")
for pname in PATHS:
sub = [r for r in rows if r["path"] == pname]
c = {v: sum(1 for r in sub if r["verdict"] == v)
for v in ("identical", "recognisable", "broken")}
ious = [f"{r['iou_vs_reference']:.2f}" for r in sub if r["font"] == "IBMPlexSansArabic"]
print(f"{pname:17}{c['identical']:>11}{c['recognisable']:>14}{c['broken']:>9} {' '.join(ious)}")
print("\n (per-case order: greeting · sentence · with_digits · mixed_latin · diacritics)")
print("\nfont coverage traps:")
seen = set()
for r in rows:
if r["missing_glyphs"] and (r["font"], r["case"]) not in seen:
seen.add((r["font"], r["case"]))
print(f" {r['font']:20}{r['case']:13} cannot render: {' '.join(r['missing_glyphs'])}")
if not seen:
print(" none")