3 ms·
Something like mixup augmentation might work here as well for training - blend two SDFs and say "this is 20% A and 80% B". Someone could also train a model tha
by lunixbochs 6y ago
Something like mixup augmentation might work here as well for training - blend two SDFs and say "this is 20% A and 80% B".
Someone could also train a model that given some glyphs of a font predicts the rest of the glyphs. Then we can do weird things like give it glyphs from multiple fonts as input to make a hybrid font.
- efskap 6y agoPersonally I'm curious about training a model to predict Cyrillic or Greek letterforms from Latin ones to boost multilingual coverage. I'm sure there are some learnable relationships there like Я <=> R or И <=> H.
- lunixbochs 6y agoVery interesting! Well, I'm playing with the concept a bit now. Currently writing a skia client to generate glyph SDFs as images so I can just use pytorch.
- wwwwewwww 6y agoI can't tell the relationship in your example. Do you mean relationship in terms of appearance or pronunciation? Я maps to "ya" sound in Russian, И is "ee" and Н is "N" in Russian and "ee" in modern Greek but in ancient Greek it's something like "e" in "bed" in American English.
- efskap 6y agoAppearance. I'm just talking about typefaces - absolutely nothing to do with pronunciation. Some fonts do not have Cyrillic glyphs. If they could be generated by an adequately trained model based on the other glyphs, then that font's multilingual coverage could expand automatically.
- wwwwewwww 6y agoI get it now. That's interesting - generating missing letterforms!
- necovek 6y agoGP likely meant the relationship in appearance: the two examples have Cyrillic glyphs closely resemble mirrored outlines of the Latin glyphs.