[patal]

Impressive work!

In the "mixed-case kerning pairs" quality testing image, I notice that the letter "j" sometimes reaches under the previous letter, like in "Fdj". Sometimes it creates a lot of space, like in "Fjo". Is there a stylistic reason for this? The Fjo spacing is the only thing that stood out to me.

Kudos

[chevyray]

Nice catch. That's a culprit of the `auto_kerning_min` property that you'll see on a lot of the fonts. this tells the auto kerner not to exceed that.

I added this parameter because I fouund that for a lot of fonts, squeezing letters together over a certain distance would just look bad, so I would set -1 or -2 as a cap.

It looks like that's just one that snuck past my notice. The word "Fjord" would look strange because of this. This is a good example of how even with the quality testing, things can get through, because I still have to visually glance over hundreds of kerning tests.

One thing that might be a nice adjustment is to have an algorithm that detects the "area" between two letters, so basically how many pixels can volumetrically fit between them, and flag ones that go over a certain threshold. I could then color those tuples as red in the sample text, basically the system marking them as "potential problems" that required an author's look.

[beardyw]

What I picked up from a lifelong typographer is that kerning should be about the area enclosed by the two letters. The aim is to make that consistent. I think that might help in this kind of case.

[patal]

Thanks for taking the time to answer. I don't understand why in the dj combination, j is able to reach under d for what looks like a kerning of about -4, when the auto_kerning_min property is set to -1 or -2, keeping Fj apart.

[zerocrates]

Maybe they just manually kerned "j" with the lowercase letters? The "j" line on the lowercase sample would jump out pretty strongly in a way the capitals-with-j don't on the mixed-case one.

[dotinvoke]

Why not just feed that information back into the algorithm itself?

[chevyray]

I definitely could, I would have to do a bit of tests to see what kinds of volumes deserved special treatment.

Usually the way I do things is I start by doing work manually. If I find that there's a common pattern in something I'm doing that could be automated, then I am able to transcribe it into the algorithm because I just follow the same steps I've been using in my head.

This wasn't a thing that actually came up a huge amount, as these glaring pairs aren't tremendously common. But they're just common enough that if I sat down and examined them, I could probably say something like "hey if 1.5 vertical lines worth of pixels are between two letters, kern this extra" or something like that.