[ft-devel] Supersampling of TrueType fonts, or: slight native hinting

[Nikolaus Waxweiler]

Hey list,
so as the new subpixel hinting code will hopefully soon be mainlined, I 
found something else to distract me from stem darkening: Keeping these 
damn TrueType fonts from snapping stems to full pixels like in the bad 
old GDI days. I'm opening a new thread here so that mails on the topic 
don't get lost elsewhere.

# What's the problem?
TrueType's explicit programming model leaves the result of hinting 
completely up to the designer. Font designers have long targeted Windows 
GDI environments with all of it's idiosyncrasies and limitations, which 
made strong distortions of the original outlines necessary. The newer 
DirectWrite gives designers much more room to archive a good compromise 
between sharpness and design fidelity, but older fonts are typically not 
updated to use the new possibilities. Even new font families like Source 
Sans Pro, Acumin and Ubuntu are still programmed to perform well in GDI 
and no effort was made to strive for higher design fidelity through 
DirectWrite. This typically means that stems are snapped to full pixel 
widths. See http://www.freetype.org/ttfautohint/#samples -> Ubuntu 
Regular, original hinting. Observe the jumping stem thickness as the 
size goes up.

This is irritating because FreeType has been using a fundamentally 
different (coverage-based) rasterization mechanism for years that makes 
all the GDI contortions unnecessary while still looking very good. See 
the above sample again. According to Dave Arnold of Adobe, the 
contributed CFF engine does not honor all Type1 hinting concepts, e.g. 
stem snapping, because it simply isn't necessary for FreeType.

This leaves us with a gigantic pile of TrueType fonts that render worse 
than they could.

# Possible solutions
Microsoft uses a 6x1 (x, y) supersampling approach to display fonts, 
which takes care of vertical stems but won't do anything for horizontal 
ones. An alternative 6x5 approach may be used at bigger sizes to 
smoothen the y-axis. Adobe uses a 6x2 approach in CoolType to smoothen 
stems to 50% coverage instead of full pixels. The new minimal subpixel 
hinting code solves the problem for the x-axis by simply ignoring all 
hints for the x-axis, but won't help for the y-axis.

I want to find a way to get fonts off their full-pixel snapping. Some 
approaches I have through of:

- Use a form of supersampling. I'm not sure how to do that. Since FT 
subdivides a full pixel into 64x64 subpixels, I tried modifying the 
rounding functions and halving CVT cut-in and minimum distance when on 
the y-axis. The results are underwhelming (outer edges of stems not 
aligned to pixel grid at various sizes) and glitchy. Another approach 
would be to double the size of a glyph before hinting and downscale 
later, although I don't know how to do that yet.

- Modify IUP[y] to go through all touched points and look for stems 
whose edges were snapped on both sides to reset the inner edge. It would 
then get picked up by IUP and follow the outer edge. This would result 
in a 'slight' native hinting mode. The downside is that this would 
probably be very difficult to generalize and would need stem-detection 
machinery.

Anybody have any ideas?