This kind of font would consist of a program or script to be used in generating each letter. Rather than every "A" looking the same, the letter would be slightly different in each iteration, with each iteration's letters designed to look good together. For example, a program could describe a slightly curvy "A" with more of a curve on top, and with a background that looks like oil paint.

The font interpreter would then generate a new iteration of that with each letter typed. Programs making use of the font would return the unique ID of letters immediately around the letter typed to the interpreter. The interpreter would then generate a letter that would have oil-paint-like strokes that line up together.

Each letter would be largely the same and easily recognizable as the letter, but have more or less of a curve in different parts, with very slightly different line widths, and a different color or image gradient in the background if so desired.

The advanced version allows arbitrary degrees and types of inputs such that the interpreter run on a compatible machine could make brighter colors if the room is bright, automatically contrast against a non-font background, or even animate based on the user's probable mood.
-- Voice, Apr 28 2021

https://en.wikipedia.org/wiki/TeX [pertinax, Apr 28 2021]

lots of fonts https://www.fontspace.com/category/unique
[xandram, Apr 28 2021]

Explorations in machine learning and latent space type design https://www.100arch...t-space-type-design
Aha, thought it was too obvious an idea not to have already been done. There's a video showing a traversal over the latent space of an S which is kind of interesting. [zen_tom, Apr 28 2021]

Another StyleGAN walk over fontspace https://twitter.com...74922050281473?s=20
This one looks quite neat - each frame returns the fontspace at a particular point in some n-dimensional space, the smooth transitions demonstrate a small step from one position in that space to another one relatively nearby. The result, an almost infinite multitude of fonts retrievable by some n-dimensional vector. [zen_tom, Apr 28 2021]

Beowolf Font https://letterror.com/fonts/beowolf.html
[xenzag, Apr 28 2021]

https://erikbern.com: analyzing-50k-fonts-using-deep-neural-networks https://erikbern.co...eural-networks.html
Source material for the twitter link [zen_tom, Apr 28 2021]

Frieze article on Beowolf https://www.frieze.com/article/letterror
[xenzag, Apr 28 2021]

Up next, combine the program with deep fakes: movies that can change voices and expressions to suit the viewer's desires.
-- Voice, Apr 28 2021


Sounds like a job for TeX (see link).
-- pertinax, Apr 28 2021


maybe I have misunderstood this idea, but I thought this was well baked. See link

Do you mean that they are never identical letters?
-- xandram, Apr 28 2021


The latent space of a GAN trained on a bunch of different fonts might be a good source of content (as well as being an interesting asset, potentially providing a single encoding model for all fonts ever). Assuming such a thing existed, you ought to be able to create a fairly rich space to traverse, either randomly, or along some path that started at whatever coordinates encoded "Comic Sans" and smoothly walked its way towards the 1750's "Baskerville" font used by Isaac Newton in his Principia Mathematica, perhaps taking a detour via Windings along the way to shake things up a bit. Any smooth path could be chosen, with the effect being a smooth transition from one font style to another.
-- zen_tom, Apr 28 2021


It would be beneficial to test it on people to make sure it did not have any effects on reading speed and reading comprehension.
-- beanangel, Apr 28 2021


Why not just use comic sans for everything?
-- pocmloc, Apr 28 2021


How far away (graphically? topologically?) from an "A" do you have to get before a person won't recognise it as an "A"? I guess that's what OCR does anyway, but I would think people are better at it (context & other assumptions...).
-- neutrinos_shadow, Apr 28 2021



random, halfbakery