Hacker News Reader

Comment by deadbabe | Hacker News Reader

What do I feel like custom font generation is one thing AI could be really good at but so far I haven’t seen anything of that sort? Seems like you could easily prompt whatever vibe you’re looking for in a font, why even bother buying commercial fonts at that point. Am I just not looking in the right places?

View on HN

Creating a CJK font seems exactly the kind of thing they're still bad at, and the results I got from just trying it on Gemini, ChatGPT, and DeepSeek pretty much confirmed that. This is like the whole "draw a clock" challenge except there are 2000+ common clocks and a bajillion more obscure ones, and regional differences like the grass radical[1]. You'd probably need to start from teaching it the principles like radicals, stroke weight, etc. as if it were a human calligrapher.

[1] https://en.wikipedia.org/wiki/Radical_140#Variant_forms

> What do I feel like custom font generation is one thing AI could be really good at

Probably because you're wrong. Anybody can make a font, making a good font is a highly under appreciated art form.

I find that a “good” font is just whatever aesthetic people find popular, regardless of legibility.

Have you seen the output of LLMs lately? It's not perfect, but I bet that with enough refinements they can create "art" just as well.

Ask the models to create vector SVGs and you'll understand how far off they are on shapes.

The latest Nano Banana is pretty good, but it's not perfect yet. And many font use cases demand perfect.

Maybe the next major update will be able to do it.

Fonts are not generated as bitmaps, anyone who doesn't see how AI can and will be good at font generation is a fool.

It wasn't long ago that we thought creativity and programming were safe from AI. Fonts are entirely within the realm of possibilities within 12 months.

What a bereft way to produce beauty. "This grey slop is good enough, isn't it?" Christ, people, let's be human again.

My man, that battle is lost.

Only if you give up.

AI fonts are nothing new. But you still need someone to manually check them and it's never one-shot. It makes no economical sense to do that instead of buying commercial fonts, which are ready-to-use immediately, when they only cost a few hundred dollars.

The price rising and licensing issue will only do one thing: pushing people to AI. Perhaps they see it as inevitable so they're trying to milk the last batch of customers though.

perhaps when the clocks at https://clocks.brianmoore.com/ consistently make sense, AI could make a font.

Even then, I wouldn't want it making a kanji font. Consider 感 and 惑, both of which would be taught before high school.

Sounds like you don't know the CJK font market. AI assisted font design is nothing new and especially useful for CJK since there are so many glyphs. There are plenty of foundries openly advertising AI-assisted fonts, e.g. https://izihun.com/fontxiazai/ziti-2673.html

Also, generating images and programs are basically orthogonal. AI could generate impeccable photorealistic images of clocks years ago, and they're much more complex than font glyphs (specifically talking about transferring a style to other glyphs; you still need to do the initial design to get something appealing, obviously*).

*Edit: Maybe AI can even handle the initial design now, not sure. What I’m saying is AI-assisted style transfer in CJK fonts is definitely old news and commercially available.

Those fonts look awful in a hard to describe way. Font uncanny-valley? I feel like a barcode reader trying to OCR meaning out of ink blots.

Those fonts are completely indistinguishable from the average cartoon/artistic font from the 2000s which are definitely not created by AI. I don’t like most of them, but I don’t like most of the human-designed ones either. Plus AI-assisted fonts in CJK often means hand drawing a couple hundred characters, maybe more, then generating the thousands of remaining characters (I assume the current crop of SOTA models could change that, but these fonts have been going around for a while), so the odds are what you see in the samples are mainly or maybe even completely human-generated.

I do not know the market well; very interesting, thank you.

Wow, K2 is killing it on this for me. Three minutes have been 100% correct now.

Haiku 3.5 had a one right too. But k2 apparently is very good at html and css.

As I understand it you can't protect the appearance of a font so imperceptible differences can be made on a copy unless you do a design patent which only lasts 15 years and isn't always done. (In the US)

It would be pretty easy to make a font generator using LLMs and visual models.

A few hours ago, [0]

[0] https://news.ycombinator.com/item?id=46127400

Perhaps an old-school bitmap font would be within the realm of possibility, but I wouldn't be confident of GenAI's ability to go from sprite sheet to proper TTF with glyphs described as curve/points without a LOT of manual work - and especially not when we're talking about a language with complicated logographs.

Reply on news.ycombinator.com