Can AI replace 99% of font design work soon?

One year ?
Six months ?

The only area I think there will be significant changes (as there already have been) is in spacing and kerning. These are areas that can be trained ā€œmechanicallyā€ to a great extent, and with enough reference material and understanding thereof, completely automated spacing and kerning don’t seem out of reach.

The creative ideation process is something I don’t see replaceable. But I will be proven wrong, surely.

An AI can’t even properly write a truly convincing creative text, so I doubt it will get a grasp on creative expression in truetype curves any time soon.

  1. Prototypo: Prototypo - Typedesign Tool
  2. Fontjoy: https://fontjoy.com/
  3. FontMap: https://github.com/kaonashi-tyc/FontMap

Doesn’t exist anymore.

What does this have to do with type design?

404/private repository.

Is there enough data out there to train AI to make fonts? They are complex, so I imagine for AI to really ā€˜understand’ type (curves, historical models, etc), it would take loads of them, which seemingly doesn’t exist in that quantity. And that not taking into account the creative part of it. There are AI generated fonts, but I don’t feel in danger yet. Famous last words?

AI is very good at imitating existing font styles instead of creating new ones.

Personally I cannot wait for AI-assisted CJK design workflow.

1 Like

CJK font design is too laborious. AI is the future and an inevitable outcome. Will it take a long time? I think Glyphs should proactively introduce AI to significantly reduce mechanical operations, allowing designers to focus their energy on the most essential creative work.

2 Likes

1 month for the .ckpt model to be ready, plus whatever time it takes to integrate it into the app. It’s pretty much solved already… the month is only for improving it (fine tunning).

2 Likes

Great.

These might be useful for you:

This is the complete CJK character structure database:

This is the font file that contains all CJK characters:

http://cheonhyeong.com/File/TH-Khaai-T4.0.0.7z

http://cheonhyeong.com/File/TH-Ming-4.0.0.7z

Maybe… but i don’t thinks so… since I should be able to handle all languages all together, past present and future ones… so there has to be nothing specific of any particular one, it must be completely general (otherwise it will fail).

Well… AFAIK… the exception will be the Inca Nots, everything else should be no problem.

Awesome !
Amazing !
Incredible !
Spectacular !
Phenomenal !

However, this preprocessed data can significantly improve the efficiency of AI processing Chinese characters. Chinese is quite distinctive, and using the general approach of AI may result in low efficiency, which is not as good as manual processing. Would you consider it?

What’s the difference of… for example… Chinese, Mayan and Latin?
As I see it, there is no difference: They are all just blobs of black and white mass.

Um, of course, you’re right.

The difference is the available training data?

To make them trainable data, some work needs to be done to connect them with AI. (The components of Chinese characters are almost always also CJK characters, and AI requires this basic knowledge to process Chinese characters.) These IDS data provide information on the components of each Chinese character and their relative positions, making it easier for AI to determine the structure of any given character. For example, when Glyphs opens a Chinese font and needs to split one or more characters into components and automatically associate them with intelligent components, it requires access to this database. The database gives AI knowledge of the structure of Chinese characters, allowing it to associate the components of many characters with intelligent components simply by linking them to the IDS.

A simpler explanation would be: If AI has access to the IDS database and knows which components make up each Chinese character, their relative positions (whether they are on the left, right, top or bottom), when it recognizes a character with a different style, it can more easily split it into the correct components according to the correct structure. Then it can associate those components with the correct intelligent components.

The structure of Chinese characters is more complex than that of letters, and IDS’s actual function is to break down Chinese characters into simpler structures. This makes it easier and more efficient for AI to recognize the shape of Chinese characters, increases accuracy, and saves a lot of machine learning processes.