I have developed a process to create typefaces using machine learning.
I trained a model on images of typefaces of different categories, and used that model to generate new typefaces within and between those categories.
The first actual font I published based on this technique I appropriately named Artificial Intelligence:
The above typeface was created by the AI – I manually made the vector and the font.
I consider this font a proof of concept more than anything else; In the majority of cases, the tech would be used to complement and assist human creativity.
The technique is, in my opinion, most powerful when used for complementation and styling of existing fonts.
Integrating this into a font creation software such as Glyphs would obviously be incredible, although I am not sure this is the most efficient approach, as it creates bitmap typefaces and also requires a powerful GPU.
My hope is that this can help other typeface designers out in their workflow. I believe it has the potential to make font creation more efficient.
I like it, and I agree that integrating this into a font editor would make sense. IMHO this is the logical continuation of other already present automatization features (both build-in, and plug-in).
I understand correctly that this is something you have only implemented on a local machine, so we can´t actually test run it, like Dall-E or so?
It should be possible to run that inside Glyphs. And combine it with outline tracing it should get you a good starting point. Is there a way to try this (with your training data and/or with someones own)?
I have not published it online, no.
I can however run requests. In its current state it is limited, however, since I trained it on seven styles only. As font makers know, you’d need many sub-categories to cover all variations.
I can also share the model through DM, which can be run locally or on a rented machine.
It should still be possible to do the process described. As for my model, I don’t know if I still have a copy of it, but that is more of an example anyway.
I am pretty sure there are more efficient ways nowadays. For example, there are web apps that are dedicated to image training, so you don’t really need to do all the backend work.
… And most importantly, the newer image models have way more training data, so to get similar results as what I show, you’d not even have to do any custom training. They already know what “blackletter” and “handwriting” is.
To begin with, anyone certainly has the right to do whatever personal experiment they want, especially with open source fonts. If I want to print a poster, or put it in a personal program, or try it in machine learning.
Theft, on the other hand, is a legally defined term. We can argue whether downloading a movie etc is theft, depending on what jurisdiction you find yourself in, but nowhere would AI in itself be theft.
Sure, I could have asked every author of the fonts if they were fine with the usage. But the fact that they published it under a license that explicitly permitted it is a pretty good hint that they’re fine with it.
And this is not even touching on the fact that typefaces don’t have copyright protection, that I did not publish the model anywhere, that I also used if for my own fonts, etc etc.
Start a new thread and tag me there, and I’ll be happy to discuss it more.
Open source font licenses require open sourcing any derivatives though. I also assume a good chunk of all these “10,000 open-source fonts” are not even open source but free trials and whatever you may find on freebies websites without reading the actual licenses.
You’re selling the AI typeface for $50.89 per style. I’m certain “Fontface” will end up working like ChatGPT, based on tokens or paid subscriptions. See, the key difference is this: designers who share their source files contribute to the community by making high quality tools accessible to anyone, while AI developers mostly just take from it.
It’s somewhat ironic that you’re trying to attack me for this, considering I have made hundreds of fonts and share everything for free for personal use.
The project (which I made in 2022 and is very dated) was just a fun test side project blending two things I like: type design and ML.
As I wrote, the font I released based on this was trained exclusively on my own fonts. You can try to twist it into being some hugely immoral thing, but I think that has more to do with your general dislike of AI than this project.
And again, start a new thread and tag me, since this one has absolutely nothing to do with the morality of AI.
I’m afraid this kind of tool is straight up ripping off other people’s work. No proof what source of data they are training with, and to be honest, even if they really trained with open-source typefaces, they should have asked the consent to the designers first, I myself would not agree with using my work for such purposes.
I totally understand that perspective. Having an AI make a derivative font based on open source and then selling it would be morally dubious. Personally I’d use such a tool only as a starting off point and would open source whatever I made. Plenty of type designers use open source fonts as the basis for their own open source fonts… this isn’t theoretically that different, unless people make a profit from selling whatever this AI generates. Potentially they could market such a tool towards designers who want more control over custom type in design projects, rather than as a tool for creating commercial typefaces.
No proof what source of data they are training with, and to be honest, even if they really trained with open-source typefaces, they should have asked the consent to the designers first, I myself would not agree with using my work for such purposes.