3 ms·

Yeah, that's what I was thinking. I wonder how big the data set would have to be. Finding a sufficient number of fonts with quality kerning might be a challen

• brudgers 11y ago
  Just off the top of my head, I think there are stages to the problem. The first is classifying the quality of kerning. For that the training set would be pages of good and bad typography. The second stage is constructing good pages from a single set of glyphs. Programs like Amazon's Kindle do this reasonably [I'm not sure they use kerning] in real time with very limited computational resources. For typography, several orders of magnitude more resources could be employed in batch mode to achieve better results. The final stage would be generalizing the ability to layout a single set of glyphs to any arbitrary set. The good news is that there are many examples for the first task and good tooling to accomplish it. For the same reasons that generalizing layout is a later step, restricting the domain to a single format such as journal articles or "For Dummies..." style books is a good way to go if possible. In other words, start with the narrow case.