Can you tell the difference between AI-generated vs human-made fonts?

ericlu · July 4

@Thomas Phinney It was clearly disclosed in the original repo of the font at the time of submission (https://github.com/mixfont/gothic-gumdrop). I never lied about the origins of the font, and I proactively stated it in the submission ticket as well.

@Igor Petrovic I'm not a lawyer myself, but I believe this is legal. My goal is to build a legal and ethical business that can be sustainable long term.

@Mark Simonson I appreciate your clarification and I do agree that the quality of some of the fonts in the quiz can be improved. However, I think it is important to acknowledge that some of the human-made fonts that I selected are being sold commercially from professional foundries that I assume do work with professional and experienced graphic designers. I personally paid for the correct licenses myself in order to legally use them in the quiz. So you have to understand that it feels like what is "good quality" to me seems like a moving target.

@Christian Thalmann I am actively here in this forum in order to engage and listen to what you all have to say. I recognize that I'm not an expert in font design and that I have a lot to learn. I am personally very motivated to improve the quality of the fonts that are being generated and I'm glad that this has been sparking discussion.

@Dave Crossland Yes I am logging the results of the quiz and I hope to share a larger write-up soon. While I can't track where all the quiz takers are coming from, one interesting data point so far is that 8% of people who completed the quiz scored 100%.

@C.Fransen I would love to discuss your ideas on how the future could be built through partnership. One idea I had is to launch a way to do a revenue share, where if you provide fonts or data to be used for training, then we could split the revenue of trained model outputs. Obviously this needs more thinking and discussion but it could be a viable path for a world where AI improves.

@Ray Larabie I really appreciate your feedback about the fonts and your suggestions for how they can be better tested and then improved. It's not my intention to create a negative and uncomfortable environment. I think that if it wasn't me, it would be someone else who invents and productizes this technology. I realize that the technology has implications but I also think that discussion and preparation for that it entails is important. I also have been finding that being "inside the cage" as you say is important in this industry - and my hope is that I can find to be a part of it and work with you all together (though I still feel that at a high level any industry controlled by a small group of insiders is subject to elitism and politics, and ultimately doesn't benefit customers).

@Simon Cozens I would be curious what your take is on a truly impressive AI-generated output. I wonder if there's a way to define that even more - what are some fonts that you feel are truly impressive and well made of the highest quality? Are there any in the display or decorative categories that you feel fit your definition of good? I would love to try to see how far the technology can be pushed right now and see if I could create a font that might match up to your standards. Obviously I'm biased but I feel like a discussion around the "state of the art" in font generation isn't complete without an analysis of the Mixfont model and also where it shines and where it fall short. I want to be a part of that conversation.

One item I'd like to note is that all of the fonts generated in the quiz were generated in less than 1 minute, and I applied no post-processing or human edits to them. I just directly generated them and downloaded the TTF to embed it in the quiz. Cost-wise its just a few cents of GPU time. I know that I can improve the quality of outputs just by turning up the compute and processing time. I also feel confident in a more expensive pipeline that can look through kerning and spacing of a given font. Would you be interested in collaborating on this? I feel that if I spent more compute budget at a specific font (maybe in the 100 dollar range) I could really improve the quality significantly. This has been my plan for Gothic Gumdrop at least for a first test.

I appreciate everyone's input and feedback. I will try to prep a write up on the quiz results which I think might be interesting to the folks here. I want to re-emphasize that I'm here to engage and learn and I would love to collaborate or partner on unique ways AI could help you in your workflows. I realize that the folks replying in this thread might not be open to this, but for those of you out there who are just reading and are curious to collaborate, my DMs are open. I would love to work together on some interesting projects like restoring old or unfinished fonts that you've started, exploring new drafts, extending your existing work, partnering together to release a joint model, or even publishing a font together - let me know.

John Hudson · July 4

So you have to understand that it feels like what is "good quality" to me seems like a moving target.

It’s really not. The font market is large, and as with virtually any other product, the quality within that market place is very mixed. So just because something is for sale, does not mean it is good quality or that the person who made it was a professional type designer. Indeed, the notion of ‘professional’ in the font business can be quite vague. There are a lot of fonts made people who have limited experience and have not really developed their skills; there are fonts made by people who are actually professionals in some other field such as graphic design or computer programming.

I suggested above some criteria by which one can recognise good quality fonts, including both technical quality and successful fulfillment of the design system.

Simon Cozens · July 4

ericlu said:

@Simon Cozens I would be curious what your take is on a truly impressive AI-generated output. I wonder if there's a way to define that even more - what are some fonts that you feel are truly impressive and well made of the highest quality? Are there any in the display or decorative categories that you feel fit your definition of good?

For heaven's sake, man. We've all told you that what you need to do is to understand fonts for yourself, and yet you insist on being spoon-fed. I shall say it one more time: You need to know for yourself what good output looks like, or else you cannot tell whether your own output is good or not.

I've even told you how to get started. I'm not going to apologise for encouraging you to learn more about the discipline you're wanting to be involved in. Just learn. Learning is good.

Obviously I'm biased but I feel like a discussion around the "state of the art" in font generation isn't complete without an analysis of the Mixfont model.

Point me to your Arxiv paper - or a complete technical description of your model - and I will be able to analyse it. Otherwise I simply can't.

C.Fransen · July 6

@C.Fransen I would love to discuss your ideas on how the future could be built through partnership. One idea I had is to launch a way to do a revenue share, where if you provide fonts or data to be used for training, then we could split the revenue of trained model outputs. Obviously this needs more thinking and discussion but it could be a viable path for a world where AI improves.

Hi @Ericlu,

I have to search for the right wording in English; if I come across as a bit blunt, it is unintentional, given that English is not my native language;

Unfortunately, I have to decline your offer for a partnership. Call me old-fashioned, but at this moment, I am not very interested in improving Ai in the area of font production. Designing type is a craft that I would like to master very much and learn as much as possible about. Handing over this opportunity would contribute to the decline of the craft, both for me and for countless other interested parties. Even for those who depend on it, because it is their livelihood. I cannot bring myself to contribute to making artisans redundant, and I even consider it downright criminal. I want to learn for myself. I read works by Morison, Updike, Lawson, Smeijers, Unger, and many others to understand and learn the craft as much as possible. I would advise you to do the same (if you haven't already). I think it will bring you a lot and provide insights to better 'understand' the profession of type designer.

Since Ai exists, I would rather see it used to solve serious problems, such as finding a cure for cancer or other diseases, stopping world hunger, or even ending wars and all the accompanying suffering. When that happens, I will be the first to take to the streets cheering. Replacing human creativity with Ai is something I have actually grown to dislike a bit along the way.

Perhaps my opinion will change in the future. That happens quite often, but for now, it is just the way it is.

Jens Kutilek · July 7

Thomas Phinney said:
Is lying about the origins of the font not a problem?

The submission checked this box:
“I am the sole copyright author of the entire project, or all other copyright authors have licensed their work to me under the OFL, and I commit to clearly disclosing if AI tools were used in the creation of this project.”

Can somebody legally claim to be the copyright author of an AI-generated font at all? IIRC copyrights have been denied on AI-generated music and images.

Igor Petrovic · July 7

@ericlu Before putting much work into it, consider informing yourself about repeated claims from professionals that this kind of ripping fonts from images is done under the loophole of "fair use". This is not fair use in its legal sense and purpose, and I am sure that it will be ruled out soon. Not to mention jurisdictions where it is already clearly illegal.

Carl Crossgrove · July 8

Eric, maybe this is not satisfying because you ask the wrong question. Most of us can definitely discern if a typeface is well-crafted. There just weren't any of them in your test. You were blurring the distinction between well-made fonts and amateurish fonts by simply comparing bad human-made fonts to bad AI fonts.

The real question, which might get you farther, could be: What quality level are my AI fonts able to achieve?

1: crummy, kludgey student work that nobody should be exposed to except a type design professor (looking at you, DaFont). A rough, un-informed stab at "making my own font".

2: Eagerly-developed but rough, shaky, advanced student fonts. Can be used in documents; they won't explode, but problems become evident with each use. Not quite finished.

3: Decent, functional fonts. Not revolutionary, not scintillating, but they work. If they offer something unique, there could be decent interest on the market.

4: Ambitious, thoroughly-expanded fonts for a special purpose. Well-tested, complete, very few details to distract or need adjustment.

5: Classics of durability, versatility, solid and reliable in every use case, dependable; nothing jumps out as a problem, flaw, or missed opportunity. Rock-solid, expertly crafted, thoroughly reviewed by many eyes. Professional grade.

Something to keep in mind: You're asking true experts and professionals in the field, whose focus has been on typefaces that work well technically and aesthetically. You have a completely separate agenda that doesn't YET appear to have any value to that pursuit: Can AI make good fonts? What if you develop your tool to a point where these experts are impressed? Because right now the AI fonts are at quality level 1 and 2.

For those of us making fonts since the 1970s, 1980s and 1990s, we obviously don't want to share or show or distribute any products that aren't at level 5, or 4 at the least. It isn't elitism driving that pursuit. It's training and experience and discernment of quality. I threw away my first 3 typeface designs. They were not at a quality level worthy of any kind of attention or value, and that just becomes more evident as the years go by.

Again: Experts can tell fine differences in quality, even if you can't. Don't make me give you the type design version of the "Cerulean" lecture from Devil Wears Prada. You don't know what you don't know.

Vasil Stanev · July 10

Nick Shinn said:

The world is burning, and the broligarchy is hyperscaling an invasive technology that consumes prodigious amounts of water and energy.
Typeface “quality” is not the issue.

You remind me of the 500+ people that I once interacted with on a certain big public Fakebook group.
First I made a post about nobody reading books because they can buy e-books online, listen to podcasts etc.. Downvoted, hiss, boo. O.K.
Then I wrote a comment about how the spread of e-books has saved tons of trees from being turned into books. Confused upvotes, confused downvotes followed.

This is how shallow most people are. In the future it will probably turn out that some data center has actually solved some majore issue in some field, like e.g. managing the logistics to a granular detail of clearing Bangladesh of toxins or recycling large tracts of the word's trash. Or sending robots to create the economic basis for an universal basic income. Then again, confused upvotes will appear. "Oh, nobody is doing actual work anymore" the people will yell. "Now robots have made us so wealthy that nobody feels the need to work anymore. I'm so fed up with having enough to eat, a mansion and two cars! The state should also provide me with a second helicopter! Down with Big Tech!!"

Nick Shinn · July 10

Shallow fellow that I am, I remind myself of Monty Python’s “What have the Romans done for us?!”

Vasil Stanev · July 10

Nick Shinn said:

Shallow fellow that I am, I remind myself of Monty Python’s “What have the Romans done for us?!”

I have no idea why you would think this about yourself, but my point is on point.

Nick Shinn · July 10

I don’t think so.
There was nothing in my statement that is inherently anti-tech-progress.

1. The world is burning. We are experiencing record heat waves and forest fires around the planet.
2. Broligarchy. American robber barons are controlling the AI process.
3. Hyperscaling. The new tech is being implemented at massive speed and scale.

Points 2 and 3 pose serious challenges to democratic governance.

4. Prodigious amounts of water and energy. Even if, as you fantasize, AI may offer some future remediation of the climate crisis, pouring fuel on the fire in the immediate present is foolish.

However, I am actually the contrarian you suppose, and have explained my philosophy in depth in this essay, Against Science—which informs the direction my type design is presently taking.

Vasil Stanev · July 11

Nick Shinn said:

I don’t think so.
There was nothing in my statement that is inherently anti-tech-progress.

1. The world is burning. We are experiencing record heat waves and forest fires around the planet.
2. Broligarchy. American robber barons are controlling the AI process.
3. Hyperscaling. The new tech is being implemented at massive speed and scale.

Points 2 and 3 pose serious challenges to democratic governance.

4. Prodigious amounts of water and energy. Even if, as you fantasize, AI may offer some future remediation of the climate crisis, pouring fuel on the fire in the immediate present is foolish.

However, I am actually the contrarian you suppose, and have explained my philosophy in depth in this essay, Against Science—which informs the direction my type design is presently taking.

Dude, can't you tell when you're being trolled?

Thanks for the manifesto link, I'll check it out when I can. Meanwhile I'm using technology to come across music to wake me up while I work:

https://www.youtube.com/watch?v=jCxgUPFVFkA To be a type designer doesn't mean I have to remain a square as I once was. I'll owe you a beer if I manage to meet you at ATypl or elsewhere

Cheers and have a smile on the house.

Edit: Checked your site, I'm thinking of buying Handsome and Pratt but this will happen later this year, probably.

No need to send discounts for now as some others have done, full price is ok.

Dave Crossland · July 11

Carl Crossgrove said:

The real question, which might get you farther, could be: What quality level are my AI fonts able to achieve?

Now there's a moving target

https://www.reddit.com/r/artificial/s/zaCodcYpjR etc

Mind the gap^H exponent

Can you tell the difference between AI-generated vs human-made fonts?

Comments

Categories