Creative Fabrica - AI Font Generator
Comments
-
I don't think describing outlines with Bézier curves is the right solution. It’s quite difficult to train a model in this way on the relatively small dataset available to us and expect good results without complex post-processing (which would distort the generated output). I believe better results can be achieved by using more sophisticated and more structured methods of describing glyph shapes.
1 -
Even if someone passes a law regulating AI training, it would be local and temporary. You can't really expect to stop the tide with legislation.Dave Crossland said:
0 -
Now I'm on a laptop not a phone I can easily find the previous thread I was thinking of.Alex Visi said:
I can imagine a drawn pdf, but for any dynamic text, such as on the web? How would search and the basic text operations work? How would it get a consistent typeface across all the branded materials? How would it go beyond generating mediocre typefaces?Dave Crossland said:
Maybe I went to far before, saying that even the Unicode text strings of the document file themselves can be discarded everywhere soon enough, but they are discarded in video data files' on screen text already since forever...
There I said the same as Yuri - for any dynamic text, such as on the web, a web page that is just a single big JPEG used to be a big problem (copy and paste didn't work, translation didn't work, search engine indexing didn't work, screen readers didn't work, etc) - but today the same A.I. technology is now very good at "computer vision" and can do all basic text operations work on such an image of a text document, by recreating Unicode strings from such images on the fly, and can even replace their regions on the image with translated texts in similar type styles.
If that sounds like more expensive computation, sure it is, but... That never stopped any of us before - even just posting on this forum is more computation than when the type community discussion was on Usenet comp.fonts. Computation of everything is always going up. I often look with much wonder at the 1970s Smalltalk desktop/network systems that faced - and addressed - most of the problems that do exist with personal computing, and arguably addressed them better than what we have 50 years later today. I mean it's kinda laughable, really, how much more powerful our hardware is after 50 years, and how less powerful our software is in comparison.... In a way, the current jump in functionality of A.I. is putting us back up on the trend line that was sagging. Anyway.
Generative A.I. systems can produce consistent typefaces and other branded materials across surfaces/sessions using a simple seed handle or hash id; these systems are an enormous mountain of data, and each prompt is rolling a little ping pong ball down, and the path it takes through the terrain can be massively compressed into a very dense ID of essentially the key turning/bouncing points. It's effectively a compression scheme, where a very large database is preloaded generically, and then each typeface's unique font ID data is a very small piece to transfer. This is all already in use for consistently recognizable human faces/appearances.
Regarding quality, Pablo Impallari already demonstrated here on this forum these systems can go beyond generating "blurry averaged" typefaces to produce "original" ones, and any mediocrity will fade away over time as more data is added and reinforcement learning recurses.
Yuri, when I saw 'laws change,' I mean the UK law has proposed to make it clear that legal challenges to this happening won't be allowed there, not to make them easier. The tide is going out inexorably.
0 -
Intuitively, I would agree that outlines only encoding in a sequential way is not enough. But, disagree that encoding outlines (bezier drawing data) in other ways e.g. graph representations, or combining with multimodal e.g. sequences and maybe outline auxiliary points for CNNs, or distance fields would not be sufficient.Yury Yarmola said:
I don't think describing outlines with Bézier curves is the right solution. It’s quite difficult to train a model in this way on the relatively small dataset available to us and expect good results without complex post-processing (which would distort the generated output). I believe better results can be achieved by using more sophisticated and more structured methods of describing glyph shapes.
My experiment showed that for the dataset, it is enough to have up to a few dozen normalised pieces per topological style. I also trained only on outline sequences, not using any additional representation, which would improve the results greatly.
The only problem I suspect is the style transfer limitation. However, augmentation of the dataset to bring some noise might help. (Haven't tested it yet.)
Again, I must stress that the training was really cheap compared to training on images.0 -
This is analogous to how AI may look at a photo of a cake to figure out how to make a cake vs a recipe.2
-
I don't quite understand what you mean Stuart, could you explain in more detail?
0 -
@Dave Crossland try looking at photos of an airplane and build one that actually flies with only this information.
That's what I'm getting at . . . AI can only generates an output to your query but it fundamentally isn't starting from any meaningful place of understanding of all the mechanical or aeronautical engineering from the ground up to come to its conclusion.
AI is only attempting to answer your question from the data it actually is working with . . . Images of airplanes.
So while AI is attempting to generate a set of harmonious letterforms to output images of set words in a particular style, it's not really coming at font design from the same perspective or with the same considerations as a font designer would.
I'm sincerely not making the case for humans here but rather as my initial comment stated, imagine trying to create a cake only from photos of a baked and frosted finished cake vs understanding baking, following a recipe, tasting the result.0 -
It's bit more complex than that.. since the training process not only uses "training" images, but "normalization" images as well.Stuart Sandler said:So while AI is attempting to generate a set of harmonious letterforms to output images of set words in a particular style, it's not really coming at font design from the same perspective or with the same considerations as a font designer would.
To put it in very simple words, "Training" images are images of what you want, and "normalization" images of what you don't want.. this way you will show both examples of good and bad type, so you can provide some sort of human teaching style guidance to the AI learning process.
The real burden is to get both sets of "training" and "normalization" images right, so the AI results can be good as one of your human students.
To summarize, you need to provide examples of not only good typography, but also bad typography, so that the AI can understand the difference between the two. (And this "good/bad examples combo" can be done on a very general level, or very specifically for each particular concept you want to teach the AI).
0 -
Does that mean that we can poison the AI by flooding the internet with broken fonts and garbage specimens? Russia and China are already doing that with disinformation. So it seems like we could dump crap on Reddit and social media to make ChatGPT think that Arial automatically spaced with Fontlab is the ultimate example of the craft.0
-
Yes, if you have 2 robots, you can train one to be Mahatma Ggandhi and the other one to be the terminator, it all depend of the training material you use to train each one.1
-
@PabloImpallari indeed you're correct on all accounts there, I was being somewhat loose with my explanation but my comment was more specific to the unique initial POV or starting approach of AI vs a human font designer even if both end up with a similar result in the end of ends which is a grouping of letterforms.
0 -
DeepVecFont is trained on 1001freefonts.com, so there is a case to be made that this is already happening.James Puckett said:Does that mean that we can poison the AI by flooding the internet with broken fonts and garbage specimens?4 -
There I said the same as Yuri - for any dynamic text, such as on the web, a web page that is just a single big JPEG used to be a big problem (copy and paste didn't work, translation didn't work, search engine indexing didn't work, screen readers didn't work, etc) - but today the same A.I. technology is now very good at "computer vision" and can do all basic text operations work on such an image of a text document, by recreating Unicode strings from such images on the fly, and can even replace their regions on the image with translated texts in similar type styles.Can you expect AI to be 100% reliable to handle all the information on the internet? So that every time you load a page it generates it without a fail? It can misread and hallucinate, so if the whole internet is JPEGs generated and read by AI back and forth, mistakes would just snowball. Like those trippy AI videos where everything feels familiar, but nothing is really there.Generative A.I. systems can produce consistent typefaces and other branded materials across surfaces/sessions using a simple seed handle or hash id; these systems are an enormous mountain of data, and each prompt is rolling a little ping pong ball down, and the path it takes through the terrain can be massively compressed into a very dense ID of essentially the key turning/bouncing points. It's effectively a compression scheme, where a very large database is preloaded generically, and then each typeface's unique font ID data is a very small piece to transfer. This is all already in use for consistently recognizable human faces/appearances.Can it though? So far, all the image generators seem to fail to maintain consistency. Recognizing similarity is much easier than generating it. AI is good with stuff that has a lot of wiggle room (such as natural forms), but so far is pretty useless when you need precision (facts, consistency, reliability). Given that they have already fed the whole internet of texts to ChatGPT, I’d expect further improvements to be marginal.
Those are the mediocre “blurry average”, though? AI feels like one of those cases with a massive gap between a prototype and a functional product.Regarding quality, Pablo Impallari already demonstrated here on this forum these systems can go beyond generating "blurry averaged" typefaces to produce "original" ones, and any mediocrity will fade away over time as more data is added and reinforcement learning recurses.0 -
I would remind everyone that AI is the cutting edge of out-of-control technological change that is wreaking havoc on civilization and the planet. Whatever benefits there may be in the latest invention, they are outweighed by the scale and speed of change. Administration can’t keep pace.
There is plenty of life left in recent tech, it doesn’t have to be continually scrapped on the altar of Progress.
That is why I am not using AI, fun though it would be.
The “adapt or perish” ultimatum should be resisted.
We are type drawers, and AI can’t draw.
As George Lois put it, “If you can’t draw you can’t create.”
Typing prompts is hand work, but it is not drawing.
1 -
DeepVecFont is an infrastructure/training architecture rather than a final model. You can train your model using that architecture on whatever dataset you want.Simon Cozens said:
DeepVecFont is trained on 1001freefonts.com, so there is a case to be made that this is already happening.James Puckett said:Does that mean that we can poison the AI by flooding the internet with broken fonts and garbage specimens?
Just read the README.
Jizhi Wangs experiment is trained on SVG VAE subsets and Founder fonts for Chinese version.0 -
I'm somewhat aware of that. In fact, it's much easier to train it in this fork, which extracts outlines from the TTFs directly instead of having to go through fontforge.1 -
In Nathan Willis' talk at ATypI (on automated spacing), he asked an interesting question about what we think automation is. First, he asked if having Glyphs link the sidebearings of one glyph to the sidebearings of another was automated spacing. A few hands went up. Then he asked if the Walter Tracy chart was automated spacing. A few more hands went up for that one.Nick Shinn said:That is why I am not using AI, fun though it would be.
The “adapt or perish” ultimatum should be resisted.
We use all kinds of tools, at all kinds of levels of automation, to aid the type design process. Some people press a little circular icon and Glyphs writes all their feature code for them. We use various scripts to automate processes over a font, setting anchors or adjusting kerning or whatever, and everyone thinks of that as normal. There is no additional karmic merit that is gained from adjusting every node and every side bearing by hand.
A few years ago I wrote a little script called Sans-O-Matic which draws the outlines for the boring "undesigned" sans cap glyphs (L, T, I, E, etc.) based on some knobs you twiddle and the metrics of your font. That's another level of automation; not AI, just some dumb stupid Python code, but I still got some pushback on that: "If you don't like drawing the outlines, why bother?" Of course you still need to draw a lot of outlines, but I think that making the design decisions and having the machine fill in the boring bits is part of this continuous spectrum of automation.
Of course, makeotf had code which automatically drew some symbol outlines and matched them to your font, and nobody worried about that. But if an AI fills out your glyphset for you? End of the world, obviously.
3 -
Thanks for clarifying - I got you now.Stuart Sandler said:try looking at photos of an airplane and build one that actually flies with only this information.
That's what I'm getting at . . . AI can only generate an output to your query but it fundamentally isn't starting from any meaningful place of understanding of all the mechanical or aeronautical engineering from the ground up to come to its conclusion.
AI is only attempting to answer your question from the data it actually is working with . . . Images of airplanes.
So while AI is attempting to generate a set of harmonious letterforms to output images of set words in a particular style, it's not really coming at font design from the same perspective or with the same considerations as a font designer would.
I'm sincerely not making the case for humans here but rather as my initial comment stated, imagine trying to create a cake only from photos of a baked and frosted finished cake vs understanding baking, following a recipe, tasting the result.
I think I can offer a refutation of this in a couple of ways though. The first is that, if we take a photo of any random 'odd' airplane and hand it to an aeronautical engineer who already knows generally how to build aircraft, they CAN then develop a series of replica which, through trial and error, will match or even surpass the original. The nature of the mega-large scale data inside the current wave of A.I. systems is that not only can a single system turn any photo into an image styled as if painted by any of the greats in the Western canon, it can also provide detailed and accurate information about cell biology, as well as aeronautical engineering. And it can cross-pollinate these wildly divergent fields of human knowledge, better than most human beings can; that's what the "performing at BA, no now MA, no now PhD, no now Post Doc" test results mean.
In this way it is indeed "not really coming at X from the same perspective or with the same considerations as human Xer would", but it produces the same or better outcomes, faster and cheaper.
You started with a 'a picture of a cake isn't a recipe to make it' metaphor... A 18 year old company founder has been in the news lately for publishing his college application letter which got him rejected from Harvard and Yale, and part of the reason the story got attention is that he claims his company, Cal AI, is making $30M/year already. What the company's app does is exactly this - you are on a diet, you take a picture of a cake, and it infers to recipe to make it, and therefore tracks the constituent macro nutrients for you, better than you could do yourself by trying to journal them manually.
Despite my skeptical posture that human type design has months left as a going concern, there will always be some humans doing it anyway... I mentioned in the other thread about the recently deceased Pope's awful gravestone carving that one of my sisters is a professional stone carver, as despite the complete automation of stone carving via machines for many decades, there remains (in the UK at least) a residual demand for artisanally hand crafted stone work. In the other recent thread about A.I. Type, I pointed out the same kind of residual demand for horse drawn carriages in some cities, as part of the tourist sector.Alex Visi said:There I said the same as Yuri - for any dynamic text, such as on the web, a web page that is just a single big JPEG used to be a big problem (copy and paste didn't work, translation didn't work, search engine indexing didn't work, screen readers didn't work, etc) - but today the same A.I. technology is now very good at "computer vision" and can do all basic text operations work on such an image of a text document, by recreating Unicode strings from such images on the fly, and can even replace their regions on the image with translated texts in similar type styles.Can you expect AI to be 100% reliable to handle all the information on the internet? So that every time you load a page it generates it without a fail? It can misread and hallucinate, so if the whole internet is JPEGs generated and read by AI back and forth, mistakes would just snowball. Like those trippy AI videos where everything feels familiar, but nothing is really there.To clarify, I'm not predicting that the legacy web turn into JPEGs. I'm saying that, similar to how the legacy web has been eclipsed by mobile apps, the new A.I. era of popular computing does no longer require the 50 years old "fonts" concept at all, and I predict that classic outcomes/use-cases of personal computers - like making flyers for a product or event - are already being achieved without fonts or encoded text at all. And from the perspective of the previous era, there are prohibitive problems with that, but those problems are already solved with the technology of the new era.
I think you are going to be disappointed, because the rate of improvement in A.I. is exponential starting in the 1950s, and the curve is starting to lilt up. Before loading this tab I was scrolling ex-twitter, and just saw this, which demonstrates the high consistency with recognizable humans in video:Generative A.I. systems can produce consistent typefaces and other branded materials across surfaces/sessions using a simple seed handle or hash id; these systems are an enormous mountain of data, and each prompt is rolling a little ping pong ball down, and the path it takes through the terrain can be massively compressed into a very dense ID of essentially the key turning/bouncing points. It's effectively a compression scheme, where a very large database is preloaded generically, and then each typeface's unique font ID data is a very small piece to transfer. This is all already in use for consistently recognizable human faces/appearances.Can it though? So far, all the image generators seem to fail to maintain consistency. Recognizing similarity is much easier than generating it. AI is good with stuff that has a lot of wiggle room (such as natural forms), but so far is pretty useless when you need precision (facts, consistency, reliability). Given that they have already fed the whole internet of texts to ChatGPT, I’d expect further improvements to be marginal.
https://x.com/runwayml/status/1917628723903463526
It seems really odd to me to see that 'real product, available today' video functionality and be like "nah can't work for black and white type"... Crossing the human 'uncanny valley' seems like a much harder problem to me, and its cracked - and done first only because there's a lot more money riding on it.
I also saw this yesterday, specifically about the lettering now possible to create with these text-to-video systems:
https://x.com/alexanderchen/status/1917635053846229089
The other (for me, VERY funny) thing I saw recently about all this, and specifically the fear/uncertainty/doubt that it is happening and will impact one's industry, is that famous silicon valley venture capitalist, and founder of Netscape, Marc Andreessen, said in an interview that A.I. will take many many jobs, but not HIS job.
https://x.com/dylanmatt/status/19175621326079427030 -
@Simon CozensBut if an AI fills out your glyphset for you? End of the world, obviously.The ethical position is not that “parts of it are OK, so I’ll do that,” but that in toto it’s wrong.
Pardon the cliché, but Mussolini made the trains run on time.
AI is a a disaster for humanity, which will not be mitigated by any efficiencies it makes possible, here and there.
Zero tolerance.3 -
Dave Crossland said:
I think you are going to be disappointed, because the rate of improvement in A.I. is exponential starting in the 1950s, and the curve is starting to lilt up. Before loading this tab I was scrolling ex-twitter, and just saw this, which demonstrates the high consistency with recognizable humans in video:
https://x.com/runwayml/status/1917628723903463526
It seems really odd to me to see that 'real product, available today' video functionality and be like "nah can't work for black and white type"... Crossing the human 'uncanny valley' seems like a much harder problem to me, and its cracked - and done first only because there's a lot more money riding on it.
I also saw this yesterday, specifically about the lettering now possible to create with these text-to-video systems:
https://x.com/alexanderchen/status/1917635053846229089That is actually aligned well with my expectations: the dataset of human faces is endless and it doesn’t require precision. At least for a demo like this, but must be trickier for a whole movie. Good fonts, or any other design, on the other hand come in singular pieces – they are the outliers from the “data”.0 -
I find most discussion of “AI” to be utterly useless, primarily because “AI” is regarded at least implicitly as One Thing, when it appears to be many unrelated things like generative adversarial networks, several different unrelated methods of language processing, computer vision, self-driving cars, massive databases of works out of copyright, massive other databases of works still in copyright, voice prompt systems that perform slightly more efficiently than a poorly trained human whose mother tongue is not the caller’s, et cetera.A font editor smart enough to add diacritics correctly to an existing design is a Good Thing that does not need a server room full of GPUs and its own nuclear reactor. A database of every existing font file scraped from pirate sources to crudely average against one another and plop out unsatifying garbage is something different. We should stop calling them both “AI.”3
-
I distinguish between Machine Learning, which I take to be a specialised model designed to do one thing and do it well, and "AI" which is generalised LLM garbage. Machine Learning applications which create pharmaceutical proteins which don't suffer from antibiotic resistance are an unmitigated human good; stupid autocomplete chatbots can get in the sea.3
-
someone should get on the line with those rubes at the new yorker's art department and let them know about a little thing called midjourney0
-
Simon Cozens said:
I'm somewhat aware of that. In fact, it's much easier to train it in this fork, which extracts outlines from the TTFs directly instead of having to go through fontforge.
Lovely
I am curious where your trial goes. 0 -
Simon Cozens said:The problem with things which will take a day to code is that you can get a project up and running faster than you can think about the legal and ethical issues.The same issue as with AI image or video generators. From their output you cannot tell what their original training data was. So what if some company without scruples trains their model on the full Monotype catalog – how are you going to prove they did?Heck, even licensing a large enough catalog to train the model on is probably not an insurmountable financial hurdle for a big enough actor, as long as their lawyers can legalese their way around EULAs to make it OK for them to use those fonts as training material. 50.000 fonts à $50 is just 2.5million.1
-
However, the specialized models are getting subsumed by the general ones.Simon Cozens said:I distinguish between Machine Learning, which I take to be a specialised model designed to do one thing and do it well, and "AI" which is generalised LLM garbage. Machine Learning applications which create pharmaceutical proteins which don't suffer from antibiotic resistance are an unmitigated human good; stupid autocomplete chatbots can get in the sea.
https://x.com/ror_fly/status/19089048686784102710 -
If any of these Ai tools are being trained on free font sites, I'm glad to have contributed to ensuring that their output quality remains poor.2
-
Would anyone explain why it is better to invest time and energy in destroying rather than creating?
By destroying, I mean polluting space by ugly typefaces, as if there were not enough of them already.
By creating, I mean contributing to a model owned by Type Designers — let's say STDM (shared type design model)
I would bet that people who like type design would rather go with that one than big tech stuff.Isn't ILT a good example that fighting against monopolies is possible?1
Categories
- All Categories
- 44 Introductions
- 3.8K Typeface Design
- 817 Font Technology
- 1.1K Technique and Theory
- 635 Type Business
- 451 Type Design Critiques
- 549 Type Design Software
- 30 Punchcutting
- 139 Lettering and Calligraphy
- 86 Technique and Theory
- 53 Lettering Critiques
- 500 Typography
- 309 History of Typography
- 117 Education
- 74 Resources
- 520 Announcements
- 84 Events
- 107 Job Postings
- 160 Type Releases
- 168 Miscellaneous News
- 271 About TypeDrawers
- 53 TypeDrawers Announcements
- 117 Suggestions and Bug Reports