Explaining myself and Mixfont, an AI font generation model

Jens Kutilek

John Savard said:

Given, though, that AI appears to allow mere ordinary mortals to... bring into existence, if not create... motion pictures of feature film length, something which, when done the Hollywood way, requires immense resources, I'm hesitant to view it as good for nothing. It seems to open up new possibilities of genuine value.

Isn't it remarkable how AI is able to solve every issue, except the ones in one's own areas of expertise?

Eko Bimantara

I've tried creating a font on the site, and it does seem to be more geared towards "quick-to-create" fonts. I'm hoping to explore more of the generated results to help speed up my font-making workflow, but there are still many obstacles and need lot of addition for it, as discussed. For example, is it possible to have additional tools for refining outlines or adjusting kerning/spacing?

This refining technicality is important because it needed to achieve specific appropriateness of the work, for example in a text font family; How do you adjust vertical metrics and set up hinting? How about consistency between styles? How about optical balance and type color? In minimum level: How to make the generated extended styles not looks like faux bold? Adjusting all these needs seems difficult using prompts; it's still more comfortable to use hands with the font editor feature.

There's also concerns about impact of quick-to-create fonts like this site's do: Some people might worry that, not only AI-generated works, but genuine works which is not involving AI, can be "visible" or "labeled" as being AI-generated.

In automated writing, its easy to make apparently "smart" writtings, but there are often certain language patterns or diction that strongly suggest they're AI-generated. If we take it raw, without editing, by someone who's not used to write and doesn't have a personal expressive language, it's not hard to guess it was created by AI, especially by someone who used to read.

This font generator also produces similar results. On the one hand, it can generate seemingly complex and well-thought-out letterforms or typographic concepts, but on the other hand it's flawed and needed a lot of refining

Dave Crossland

I should point out that Andrés Torresi is the primary author of the ht Letterspacer project, and he'll present it at the Typographics 2027 TypeLab next week - it's one of those attempts Simon Cozens mentioned. 

I also want to say that I think it's nice Eric bothered to engage this forum...

The lack of reckoning with the current and rapidly developing situation at Western type events ha struck me as peculiar. I think it's important that this topic is discussed. 

First I should preface that... In https://ethanmarcotte.com/wrote/all-tomorrows-parties/ (the designer famous for cooking and promoting the idea of "responsive web design", iiuc) says that the phrase "what will happen next?" lacks agency, but per "Autonomous Technology: Technics-out-of-Control as a Theme in Political Thought" from Langdon Winner in 1977, it's unclear to me we-human-society have any. 

Back to my point... I believe the majority of computer scientists who developed typefaces generators don't bother; I expect their number to increase exponentially upwards as the difficulty of doing this goes exponentially downwards; and I expect those two trends to continue such that the next 10, 100 Eric's won't post here at all.

Impallari appears to have been the first. Eric's become second. I hope there'll be a third. 

Ray Larabie

I just want to point out that AI ethics side, all the AI generated typeface results I've seen so far are totally ass. Sure it can do the easy letters...no kidding! Any fool can extrapolate those in their sleep. Look how it botched up the numerals and other hard glyphs. That kind of problem solving is what typeface design is all about, and AI currently doesn't even come close to handling it. Wake me up when it generates something remotely innovative.

Carl Crossgrove

Eric, Simon has said something that is a bit of a mantra for me with typeface design, and a couple examples may help: Objects that also require a bizarrely microscopic level of detail awareness: Eyeglass frames, and Automobile bodies. There are likely many others. Compare the eleventy-something midsize family SUVs available, or the shape of the category of eyeglass frames popular for a long time known as "rectangular" - Or any of the sub-categories. Peruse the literal hundreds of options.

What exactly are the differences? When we try on several pairs, the differences "show up" when combined with a face, but how could one of dozens of similar pairs be so wrong, so awkward, when another pair, seemingly "the same," will pair perfectly with one human's face shape, eye positions, forehead, etc. The answer is the same with tiny, "imperceptible" flaws in typefaces: Your eyes have to be trained to SEE finer and finer differences in shape. The top of the lenses might angle down just slightly in one pair, but be perfectly level in another. The two lenses might be the tiniest bit smaller in the vertical direction but bigger horizontally. The bridge length, the frame thickness, the sharpness of curves, the lens size, every variable you can describe affects the appearance of an eyeglass frame and a typeface. This skill is necessary for those tasks, mentioned above, to be meaningfully and successfully accomplished.

I don't want this to sound exclusionary or elitist. I think people in other fields with this kind of trained perception would agree: if you can't see (or hear) the difference, how would you improve it? Because some type designers have shaky hands (me), or very unreliable motor control, they may not be able to "draw" the exact shapes they envision. Bezier [and other outline] tools have absolutely helped me to get control of shapes to a satisfactory degree. Technology has always been blended into the creative parts of type design or type making, so you can trust that probably nobody on this thread or this forum is threatened by "technology". Having the wonderful assistance of perfectly-smooth bezier shapes allowed me to take the steps of building a character set, adding different weights, italics, and then printing them out.

With a lot of guidance from my mentors, over time, with a lot of testing and looking and refining, I learned to SEE things that are nearly imperceptible. THIS is the most important, consequential skill in type design. The ability to really see finer and finer and finer levels of detail does not happen quickly. It requires a fair amount of training and critique, which I was lucky to get. Success in typeface design is primarily learning to see. 

Nick Shinn

Ultimately, the difference between text commands that create visual images (i.e. gen-AI prompts) and shapes made directly with a pen (whether shaky or not) is that such line-drawing engages more areas of the brain than typing. This was discovered in recent research. Presumably this is the result of the constant feedback loop of hand-eye co-ordination involved. 

However, if that were to result in more sophisticated designs, as one might expect, would they even be appreciated by a populace that was keyboard-centric? At least learning to read and “write” (make visible text) still incorporate hand-writing, putting all readers on the same page, as it were.