Learn how to use the forum

GPT4 integration

Discussion started on FontLab 8

I know I'm jumping the gun on this, but everyone else is doing it.

What are some ways incorporating AI could be useful? The low hanging fruit is character generation, where the character set could be uploaded and you could get a set of AI generated characters for the current glyph, which could be vectorized and placed onto the design space.

Blender is already getting GPT4 integrations:

So I asked GPT4.


Glyph prediction: GPT-4 could be used to predict appropriate glyphs for missing or incomplete characters in a typeface based on the style and characteristics of existing glyphs.
Kerning and spacing optimization: GPT-4 could analyze existing letter spacing and kerning pairs and suggest adjustments for better readability and visual balance.
Font style transfer: DALL-E2 could be utilized to transfer a specific style from one font to another, helping designers create new typefaces quickly.
Font family expansion: DALL-E2 could generate variations of a given font by creating bolder, lighter, or slanted versions based on the designer's input.
Language support expansion: GPT-4 could be used to generate glyphs for additional languages or scripts, expanding the typeface's usability across different regions.
Logo and icon generation: DALL-E2 could be employed to create custom logos or icons based on the typeface, enhancing branding and design possibilities.
Font pairing suggestions: GPT-4 could analyze and suggest font pairings based on the typeface's characteristics and context, helping designers choose complementary fonts. *tried this and currently it's nonsense
Font classification and tagging: GPT-4 could automatically classify and tag fonts based on their features and styles, making it easier for designers to find and select fonts.
Interactive font design: GPT-4 could be used to create an interactive design environment where designers provide natural language input to modify or create fonts.
Design feedback and suggestions: GPT-4 could analyze a typeface and provide feedback, suggesting improvements or identifying potential issues.



#1 - 2023-03-25, 12:42
« Last Edit: 2023-03-25, 12:45 by insigne »


I would like to propose the integration of artificial intelligence into font design applications. As previously discussed on TypeDrawers, one drawback of AI in design is that type designers may be among the last to benefit from it due to the niche nature of our tools.

I believe AI's role in type design could evolve in a different direction. Rather than incorporating AI into type design tools, we might see type design extensions added to AI platforms. For instance, with the recent announcement that ChatGPT will have extensions for mathematics, web browsing, video editing, and more, we can envision a near-future scenario where users can simply drag and drop a UFO file into ChatGPT or a similar tool. This would allow them to input a prompt like: "Add accents with anchors to support all common Latin-based languages. Include current currency symbols. Scale and adjust vertical metrics to ensure accents are not cut off."

In my opinion, this approach is more likely to materialize sooner than integrating AI extensions into existing type design tools. But if it happens: cool!
#2 - 2023-03-25, 18:14
Raymond Larabie
Typodermic Fonts Inc.

Hi,

The AI does not exist. Only human have some form of intelligence, some animals too.

AI is a big hoax, at best a good marketing

Any machine, including the bigest computer, produces only what the man who built it wants it to produce.  Then create or expand a font !!!

GPT4 produces something from what it already knows, without knowing its meaning. So creates something new that has its own meaning, something that produces an intention!!!!!

There are a lot of books explaining what is realy AI. Artificial, yes, Intellignece, no.

The best things to do is to stop the discussion here.

Thanks

A. Philippe



#3 - 2023-03-27, 09:42


The term “Artificial Intelligence” (AI) can indeed be misleading, as it is often confused with Artificial General Intelligence (AGI). It is essential to understand that GPT-4, a large language model, falls under the category of narrow AI, meaning it is specialized in specific tasks rather than possessing human-like intelligence, which is the realm of AGI.

In the context of font design, it is accurate to say that GPT-4 may not be capable of generating entirely new and creative font ideas as a human designer would. However, it is essential to recognize that GPT-4 can be highly adept at certain technical tasks and interpolation. Jeremy's post suggests utilizing GPT-4 for these specific purposes, rather than expecting it to exhibit human-like creativity.

The potential for integrating GPT-4 into font design software is an exciting prospect, as demonstrated by the Blender demo. By providing an interface for GPT-4 to work with software, it can be a valuable tool for accomplishing technical tasks and interpolation in font design. While it is crucial to acknowledge the limitations of narrow AI, it is equally important to recognize its strengths and the potential benefits it can bring to various industries, including font design.
#4 - 2023-03-27, 16:22
Raymond Larabie
Typodermic Fonts Inc.

This would allow them to input a prompt like: "Add accents with anchors to support all common Latin-based languages. Include current currency symbols. Scale and adjust vertical metrics to ensure accents are not cut off."
What is the meaning ot this "Add accents..."? What a poor demand!
Why not "Create an original Art Deco font with all needed characteristics (accents, scurrency symbols, etc.) and put my name on the copyright."?

Some forum member answered this topic saying this was frightening. His answer was deleted (I think he did it himself). I second his answer and ALPH's opinion, even if you're free to carry on.

For me, this is crap and FL8 forum is not the right place for it. Leave it to the Typedrawers forum where you can discuss it with "specialists" at greater lenght.
#5 - 2023-03-28, 06:54
I'm definitely not an expert at all ! Senior refers only to my age ! 😪

Bonsoir,

Merci de votre réponse qui est de bon sens. Les quelques professionnels concepteurs de polices de caractères que j'ai rencontré sont des artistes et des créateurs et non des plagiaires de ce qui à deja été fait. Et ils utilisent FontLab.

Le forum devrait plutôt se préoccuper de la documentation utilisateur : cela va faire un an que nous avons un nouveau logiciel et pas de documentation utilisateur. Que fait le personnel de FontLab ? Creer de nouvelles fonctions ? Cela ne semble pas être le cas car il n'y a pas de nouvelle version. Ecrire la documentation ? Dans ce cas, merci des informations?

Une documentation utilisateur complète de la V8 éviterait beaucoup de questions sur ce forum. Ce sont bien souvent de simples questions de mise en œuvre de FontLab et non des cas compliquées à résoudre.

Bonne soirée.

Salutations


Hi

Thank you for your common sense response. The few professional typeface designers I have met are artists and creators and not plagiarists of what has already been done. And they use FontLab.

The forum should rather concern itself with the user documentation: it will be a year since we have a new software and no user documentation. What do FontLab staff do? Create new functions? This doesn't seem to be the case as there is no new version. Write documentation? In that case, thanks for the information?

Full V8 user documentation would save a lot of questions on this forum. These are often simple FontLab implementation questions and not complicated cases to solve.

Have a nice evening

Thanks


A. PHILIPPE
#6 - 2023-03-28, 10:41

Quote
Some forum member answered this topic saying this was frightening. His answer was deleted (I think he did it himself).
 
Indeed, Yves, I deleted my earlier comment
 
“Doesn’t that sound like a nightmare? … ;)
 
because it seemed a little flippant concerning the vast change in creative processes we are going to see over the next years, due to the development of systems like GPT-4. And Jeremy really didn’t name just frightening aspects of integrating systems like GPT-4 into type design applications. For instance, I believe that a well-trained, parameterized kerning model might provide better starting points for kerning than some of the currently available tools. Simon Cozens had tried to set up such a model some time ago, with mixed results. On the other hand, there are so many questions connected with the use of AI applications in visual design (copyright questions being not the least important among them) that it would require a lot of thought and argumentation to come to a fair and nuanced assessment. Personally, I have yet to see results of AI applications in the area of font design that would make me worry for the existence of the profession.

In general, I would agree with critics of the term “artificial intelligence,” and I would also oppose the use of the term “model” in connection with systems like GPT-4. To my mind, these terms give an overall false impression of what the respective systems are and how they work. As a minimum terminological correction, I would maintain that “artificial intelligence” should be replaced by a term like “simulated intelligence,” for what these systems do is actually simulating intelligence without providing a true model-based account of the workings of “intelligence” as we tend to ascribe it to ourselves as human beings.
 
Take the so-called “language models” which are used by applications like ChatGPT. On closer examination, they cannot really be called “models” in the sense in which this term is used in science. A true language model would have to explain, amongst other things, how language can be acquired by human beings, and it would have to consider what Chomsky called “the poverty of the stimulus” in first language acquisition. Unlike children, systems like ChatGPT require an abundance of interpreted linguistic stimuli before they can even start with their language simulation work. A child, however, does not acquire its ability to talk based on statistics. So a child and a chatbot are very unlike systems by design, and the latter doesn’t really explain a lot, if anything, about the stunning linguistic abilities the former will eventually acquire. So in my opinion, the term “model” used for the data sets on which the abilities of chatbots are based is misleading and creates a false impression of such systems. But I don’t want to go completely off-topic … ;)
#7 - 2023-03-28, 11:44
« Last Edit: 2023-03-28, 11:46 by Alex_H »

If we would just stop to think about the practicality of it, as it applies to type and type design, do the developers of font editors really have any incentive whatsoever to add full implementations of AI (or whatever you want to call it) to their programs?





Besides, the software already does so much of the work (actually most of it), and we have to consider when we make these requests for "machine intelligence" additions: at which point would we have to stop reasonably calling the work our own?







Do we really just want to put a blueprint in the box and get it to do everything? More to the point, I find it amusing (if a bit disconcerting) that we're even entertaining the idea of a box that does everything in our field. Imagine what your own reaction would be to such requests if you were the box maker.

#8 - 2023-03-28, 12:59

Very well said, Patrick. — There were also some good points made by John Hudson on Typedrawers regarding the differences between the process of developing and refining a design idea on the one hand and the kind of statistical predictions that are made by a visual AI system on the other hand. I would also maintain that there is a fundamental difference between these two processes, not only because the first process involves “gradually refining a design in reference to itself,” as John put it, but also because this process is inherently visual and requires a sort of visual thinking that cannot not adequately captured by linguistic expressions which can be sensibly used as prompts.
 
As we know, AI systems are already pretty skilled in building things that can be captured by descriptive language. You can ask such systems to “draw an elephant with a red hat” or maybe even “build an alphabet of the Fleischmann type, but with the proportions of …” (insert a descriptive term or a name of a different typeface here), but how can you really describe what happens when you see that you need to move these nodes and those handles in this and that particular way to make a letterform “fall in place” or “start to rhyme” in the context of a word shape?
 
Sure, you can say, “move those nodes and handles in this and that way,” but then you are not using the AI, but just another drawing tool. So how do you communicate in general descriptive terms to the system what happens when a letterform “falls in place” or “starts to rhyme with other letterforms” or “begins to look right” in some context or “starts to make sense” in relation to a design idea that, in turn, becomes clearer by the changes themselves? Why would you even bother searching for words when you can do this better and in a more immediate way yourself, without using any words?
 
Put differently, I would maintain that descriptive language is either very precise (“Move node X by 3 font units to the right”), but then it is not very useful for formulating input prompts to an AI system, or it is too general to describe the outcomes of the particular kind of iterative, visual thinking that is the essence of the designer’s process in refining a design idea.
#9 - 2023-03-28, 14:50
« Last Edit: 2023-03-28, 14:58 by Alex_H »

To be fair, quite a bit of what we do is parametric, and machines are pretty good at parametric. My discomfort with the current enthusiasm for all this is at our tendency to get carried away with things — enough to sacrifice craft and visual objectivity for the sake of convenience, which is something we've historically done over and over in the name of technological progress.



I think the algorithms will eventually be sophisticated enough to figure out every aspect of what we do and probably surpass us, but I believe we're nowhere near that at this point in time. It'll be amazing (and I'll be a very old man) when a machine spits out something like Antique Olive Nord with minimal human input.



I suppose speeding up the building process is the main reason people want the advanced machine intelligence in type design programs. The most obvious way to do that would be to make the software something you can talk to — meaning never mind the mouse and keyboard, and just tell the machine what you want to do. "Reduce left sidebearing by 23 units". "Change the top left line to a curve." "Bring the O a bit closer to the V." That kind of stuff.



But then we would probably have to ask Yuri to include custom speech shortcuts in the program.

#10 - 2023-03-28, 15:29
« Last Edit: 2023-03-28, 16:11 by Patrick Griffin »

To be fair, quite a bit of what we do is parametric, and machines are pretty good at parametric.
This is true, of course. In the above argumentation, I meant to highlight an aspect of the design process that is not really accessible to being supported by AI in its current form, as I believe. And yet, many people seem to hope that AI would help them just at this point. I do not deny that AI can be useful in what you call the parametric aspects of the design process. I do not even deny that AI might be able to deliver starting points for design tasks or to create variations that can serve as exploratory means.
 
But imagine you have a new and very specific design idea for a typeface that you think you were able realize across your glyph set to a certain degree, but you are not quite content with the outcome of your attempts at catching it. You feel that your idea doesn’t come out as clearly in your glyph and word shapes as it could. Then I’m not sure how an AI system (which, by its setup, is trained on existing type designs) could help you to bring out what is missing in the design at hand. And I’m not sure what words you could use to tell the system to bring out what’s missing.
#11 - 2023-03-28, 23:39
« Last Edit: 2023-03-28, 23:41 by Alex_H »

I suppose speeding up the building process is the main reason people want the advanced machine intelligence in type design programs. The most obvious way to do that would be to make the software something you can talk to — meaning never mind the mouse and keyboard, and just tell the machine what you want to do. "Reduce left sidebearing by 23 units". "Change the top left line to a curve." "Bring the O a bit closer to the V." That kind of stuff.
But this would be, as you say, a kind of AI that performs tasks similar to what assistants like Siri already do: “Set the alarm to 7:00 a.m.” Which is something different from using AI as an agent in design. It’s what I tried to describe as using AI as a different kind of Pen or Contour Tool, so to say.
#12 - 2023-03-28, 23:47

But imagine you have a new and very specific design idea for a typeface that you think you were able realize across your glyph set to a certain degree, but you are not quite content with the outcome of your attempts at catching it. You feel that your idea doesn’t come out as clearly in your glyph and word shapes as it could. Then I’m not sure how an AI system (which, by its setup, is trained on existing type designs) could help you to bring out what is missing in the design at hand. And I’m not sure what words you could use to tell the system to bring out what’s missing.
Conversely, you would make up the words to use. Say deco gets defined by the machine as a set of parameters that include for instance a 4.8:1 width ratio, a 70/30 contrast, bars that sit at 68%, etc. Then you can tell the algorithm to "take the design a bit in the deco direction but hold off on raising the bars" — so essentially interact with it back and forth using predefined settings until you reach a satisfactory situation. It wouldn't be that much different from how GPTx works now, where you tell it something like "Write my physics homework in a dozen paragraphs using the style of Richard Feynman for two-thirds of it and the style of Shirley Jackson for the remaining third, and present it in a Paul Rand aesthetic." This of course relies on trusting that if the machine could learn to emulate Feynman, Jackson and Rand, it can probably learn what deco is if given enough data.


But this being a visual design endeavour, defining the words to use (or which data to feed into which word/category) is where it would be very difficult to avoid inherent bias in the algorithm. Your idea of deco may be different than mine and probably different from what the machine learned from the data it was fed. And who's to say that the machine's definition is less correct than yours or mine? After all, it learned through a lot more data than you or I could absorb in three lifetimes. So if the machine says deco is this and that, the vast majority of people are just going to trust it and go with it until it becomes the standard definition used everywhere.


Having ranted all that there, I think what people really want when they ask for AI in type design is predefined but modifiable algorithms that they can interact with to speed up their process. I don't think people really want the machine to do the type design for them, because the creativity and ownership of the product starts seriously becoming suspect there. A basic form of the predefined-but-modifiable algo already exists in font editors — through stacked actions and scripts. We can just learn to use them in more creative ways, or I guess ask developers to teach the machine to do that for us. That depends on our level of comfort with how much of the work we want to call ours.
#13 - 2023-03-29, 07:52
« Last Edit: 2023-03-29, 07:54 by Patrick Griffin »

But this being a visual design endeavour, defining the words to use (or which data to feed into which word/category) is where it would be very difficult to avoid inherent bias in the algorithm. Your idea of deco may be different than mine and probably different from what the machine learned from the data it was fed. And who's to say that the machine's definition is less correct than yours or mine?
Indeed, this is a crucial problem. Just think of the more general question of typeface classification which is implied in your “deco” example. Typeface designers, typographers, and historians of type have not even agreed on the most suitable form of a classification system for typefaces. They have not even agreed on an acceptable methodological framework for such an endeavor. But if there is no consensus about the proper treatment of such fundamental questions within the community, any attempt at defining a term like “deco” cannot make any claim on intersubjective validity. On closer consideration, such an attempt cannot even be described as “biased,” as this word seems to indicate a deviation from an accepted standard of objectivity.

So if you say that the designer would have to “make up the words to use” in talking to an AI system, this would entail, to my understanding, that the designer would have to train the system themselves in order to teach it a language that would be useful for the intended way of employing the system later.

For if the designer just accepted what was trained into the system by others, all the consequences described above would follow. If the designer then asked the system to “take the design a bit more in the ‘deco’ direction but to hold off raising the bars,” they would fall prey to the inevitable bias or arbitrariness that went into training the term “deco” into the system. As a result, the prompt in question, if successful, would take the designer’s work a bit more in the direction of what the other, unknown trainers had considered to be a “deco” typeface, but it would not necessarily take it in the direction of what the designer him- or herself considers to be “deco.”

Therefore, I would conclude that “making up the words to use with a system” is just a different term for training the system. But then, why would you go though the tedious process of training an AI system to understand words in a particular way, if your experience and your visual training allows you to make the design changes you intend directly? Put differently, if you have a concept of a “deco” typeface, why bother yourself with teaching a machine this concept, when you can apply this concept directly? All the more, when the concept, as your example goes, is parametric in nature and can therefore be applied to an entire font at once?
#14 - 2023-03-29, 14:33
« Last Edit: 2023-03-29, 14:37 by Alex_H »


Quote
So if you say that the designer would have to “make up the words to use” in talking to an AI system, this would entail, to my understanding, that the designer would have to train the system themselves in order to teach it a language that would be useful for the intended way of employing the system later.

Yes.

Most examples of machine learning that are currently making a splash in various media are using the vast and broad input of basically the entire Internet for both the training and the source material from which derivative output is created. The machine’s understanding of the words and syntax of the linguistic cues and how these relate to existing works—textual, visual, musical—are informed by the content and organisation of the Internet (and hence subject to the same kinds of biases as e.g. search engine algorithms). Much more interesting, to me at least, is narrow training on very specific input, because I am not interested in having a machine that creates pastiches typefaces that it finds on the Web, but I am very interested in having a machine that can assist me in creating my typefaces. That means training the machine on my types.

#15 - 2023-03-29, 15:10

This is an interesting approach, John! So you would train the system on your types in order to use it later for helping you, e.g., extend the typeface at hand and cover new scripts and languages. Or at least get you some interesting starting points for further work in these areas.
#16 - 2023-03-29, 15:16

Regarding the debate surrounding the term "AI," I believe some individuals may be purposefully misunderstanding the matter. The term "AI" is frequently employed in English to describe automation tools such as GPT.

If you genuinely believe that automating certain aspects of type design is detrimental to the craft, it seems peculiar for you to express this sentiment on a FontLab 8 forum, rather than discussing it within the context of Ikarus or Fontographer.

There are numerous non-creative tasks in type design that could greatly benefit from automation. Consider a legacy typeface with 12 styles that lacks combining accents. The process of adding those accents, anchors, and under accents can be time-consuming and laborious, with little room for creativity. The only creative aspect might be determining the height of the under accents. With programming skills and patience, it may be possible to automate tasks like this, leaving more time for designers to focus on the creative aspects of their work.

Some designers may enjoy drawing the alphabet, while others may find programming more engaging. For those who appreciate the creative side of type design, automating mundane tasks like filling the PANOSE table or blending intermediate kerning can be highly beneficial. As customer expectations rise and distributors demand more, it is crucial for type designers to optimize their workflow. While AI should not take over the creative components of type design, it has the potential to significantly streamline the more tedious aspects.
#17 - 2023-03-29, 17:39
Raymond Larabie
Typodermic Fonts Inc.

Regarding the debate surrounding the term "AI," I believe some individuals may be purposefully misunderstanding the matter. The term "AI" is frequently employed in English to describe automation tools such as GPT.

If you genuinely believe that automating certain aspects of type design is detrimental to the craft, it seems peculiar for you to express this sentiment on a FontLab 8 forum, rather than discussing it within the context of Ikarus or Fontographer.
Automating the kind of tedious tasks you describe, Ray, should be pretty non-controversial. We use automation tools all the time, and they could only benefit from becoming smarter. So that really doesn’t warrant a discussion, does it? The question of employing AI for design purposes becomes interesting only at the point where actual design work comes into play. There may be a somewhat blurry line between what is just a “tedious task” and what can be considered creative work in the proper sense, but in general, I think there is a consensus that nobody wants to go back to the days of Fontographer. So I don’t think it is a purposeful misunderstanding of the term “AI,” when we discuss the impact of the technology on creative work proper. Rather, I believe this is a question designers will inevitably have to face.

Going back to what John said, I think it will be interesting to see whether AI systems will prove efficient enough for tasks like extending the language support of typefaces by suggesting the required glyphs. Everything depends on the amount of training such system will need in order to come to initially acceptable results. If you already have a considerably large typeface like The Brill in version 2, I can imagine that the stimulus (the training data) for an AI system might be sufficient to quickly come up with an initially acceptable glyph shape for a character like U+046E that was now manually drawn for the later versions, at least partly:

U+046E.pngGPT4 integration


But if you don’t start from a fairly extended set of glyphs as included in The Brill or if you are asked to extend a typeface to a script for which no immediate shape typology is available in the already existing glyphs, it could well be the case that the AI system will suggest something to you that resembles the experiments Eric Gill once made for a Greek extension of Perpetua:
 
Perpetua-Greek.jpgGPT4 integration


I doubt this will be widely accepted as a Greek typeface. The shape logic of Greek and Latin glyphs is just different and rests on different traditions. So in order to teach an AI system how to make an extension of a typeface when the stimulus (the training data) is poor (a situation which resembles first language acquisition in humans), I suspect you will have to spend so much time on training the system that employing the AI will soon become inefficient compared to actually drawing the glyphs yourself.
#18 - 2023-03-29, 23:50
« Last Edit: 2023-03-29, 23:51 by Alex_H »

Hi,
:
I remember, that there was once a Typostammtisch with David Brezina, who wrote his Ph D research about machine learning and neuronal networks to generate accented glyphs out of a standard set of latin glyphs. He let the network go through other typefaces before to generate the necessary glyph variations of the standard set. It seemed a good technique to generate glyphs, which were not included in the font.
But this was not based on the vector input, but on a black and white matrix. Not far form the research of Adrian Frutiger go get the standard skeleton of a letter a, for example. He overlayed Garamond, Baskerville, Bodoni Excelsior, Times, Palatino, Optima and Helvetica.
As typedesigner I am happy of every support I can get, to refine typefaces. This gives me more time to ensure proper quality in form, spacing and kerning. Thank you for your interesting input.
:
Best Jan

:
David Brezina Typostammtisch Berlin
:
Ein_Leben_für die_Schrift_Adrian_Frutiger.jpgGPT4 integration
 
:
#19 - 2023-03-30, 03:51

One of the points I was trying to make with my "deco" example is that whatever the machine learns and widely repeats will most likely become a standard that the general public takes to be true, and in a specialty design field like ours this can easily turn into a madhouse. Automation has been here for a while now, and we all exploit it and want more of it. But once it crosses over into the creative side and is generally available, the specialty itself becomes a machine one, and arguing against its decisions would become the domain of Luddites and paranoid crackpots — kind of like how I'm seeming now in this rabbit hole. Why spend an hour to decide how wide that H should be if the machine can instantly tell you with the distilled certainty of history's glorious entirety that 712 units and 82.4 sidebearings is the definite bullseye? Shrugs and eye-rolls all over, and here's your scrip on the inside of a tinfoil hat.




At any rate, the kind of machine training people are asking for needs massive amounts of data, so doing it with a few hundred faces is unlikely to yield great results. If we put tens of thousands of faces in the bowl, though, we may end up with the ability to ask the machine to shake the perfect combination of spices into the sauce. Can't wait.




















#20 - 2023-03-30, 06:06


Quote
Automating the kind of tedious tasks you describe, Ray, should be pretty non-controversial. We use automation tools all the time, and they could only benefit from becoming smarter. So that really doesn’t warrant a discussion, does it?


The topic certainly merits a thoughtful discussion. While I understand that the prospect of using AI to generate alphabets may be of little interest to some, it is important to recognize that the current tools at our disposal don't allow for the extrapolation of characters. Instead, we should focus on the potential for a FontLab-AI interface that could automate some of the more tedious, time-consuming, and less creative tasks in the design process.

I kindly request that we keep this thread focused on the practical applications of such technology, rather than venturing into speculative territory. It is true that machines may generate fonts in the future, but it is unlikely that this would involve FontLab. Let's concentrate on the practical possibilities at hand.
#21 - 2023-03-30, 16:49
Raymond Larabie
Typodermic Fonts Inc.

When I re-read the wording I used in the paragraph you cited, I see that these sentences may have come across somewhat offensive, especially to a the native speaker of English (which I’m not). I’m sorry, Ray, I didn’t mean to sound harsh, and of course, I didn’t mean to brush aside the idea of discussing the possibilities for automating specific tasks with the help of AI.

Here’s something I learned in a client meeting today. The client is developing UI systems that involve a family of fonts we developed, and in order to check how the fonts perform within bounding box restrictions, they asked Chat-GPT to come up with phrases in Arabic and Thai that would involve real-world extreme cases of diacritics or tone marks placement above or below base letters or in relation to other marks. And what shall I say? — The AI did a rather good job in spitting out phrases that would showcase just what the client wanted to check in their layout grid.
 
So if we talk about practical aspects of using AI systems for type development, this is something I would definitely put on the list: using Chat-GPT for spitting out text examples tailored to specific evaluation purposes.
 
I imagine that our use of texts in the font development process can only benefit from using AI systems. Why not ask Char-GPT to come up with language-specific lists of kerning pairs, ordered by frequency in existing typeset texts? — If the results delivered by AI systems are reasonably reliable in this area, I think it would be highly useful to have an AI system for generating example texts inside FontLab.
#22 - 2023-03-31, 11:20
« Last Edit: 2023-03-31, 15:42 by Alex_H »

That's an intriguing perspective, Alex. Pretty cool! I must admit that I hadn't considered using AI to generate data for that kind of use. Maybe it can be used as worst-case Latin accent collisions or to assist with kern lists for unique letterforms. For instance, in a typeface with a rounded left side of the “F,” AI could potentially generate a comprehensive list of all plausible pairs. I'll try it next time.

Although I'm not well-versed in programming, I wonder if there's a possibility of using AI to directly edit a font's source code. FontLab 8 has a JSON format that might be editable by an AI. While it may not be practical to modify letterforms directly within the code, it's conceivable that some parameters could be adjusted. However, I'm uncertain about the specific features available in the JSON format.

By not relying solely on font software developers to add features, AI could potentially be trained on JSON or XML files for existing fonts and learn how to program them. I want to clarify that my knowledge of programming is quite limited, so I apologize if my assumptions about the accessibility of certain font elements in JSON code are incorrect.
#23 - 2023-03-31, 17:53
Raymond Larabie
Typodermic Fonts Inc.

Maybe it can be used as worst-case Latin accent collisions or to assist with kern lists for unique letterforms. For instance, in a typeface with a rounded left side of the “F,” AI could potentially generate a comprehensive list of all plausible pairs.
Some time in the past, I had developed a script that would detect collisions in glyph pairs, but the problem was that this script used to report a lot of pairs that were highly unlikely to ever occur in actual texts. Excluding cross-script pairs or lowercase-uppercase letter combinations and the like made some difference on the number of results, but there were still many pairs that couldn’t be found by manual checks in relevant language-specific text corpora, as far as these corpora were available.

Now, if FontLab provided a way to run Chat-GPT functions from the Scripting Panel, script developers could make the AI automatically check reports of such diagnosis scripts against huge text corpora. That would open a new world of possibilities for preflighting fonts in real-world scenarios.
#24 - 2023-04-02, 15:10
« Last Edit: 2023-04-02, 15:12 by Alex_H »

@Ray Larabie

I read your latest post on Typedrawers about the AI/Font creating subject. 
I followed very skeptically this thread from the beginning and must say you ended with the same conclusion as many of us about the perspective of "improved tools for working on fonts"
#25 - 2023-04-05, 06:34
I'm definitely not an expert at all ! Senior refers only to my age ! 😪

Yes. I was quite optimistic about these new tools, as they've been truly transformative in other aspects of my work. I had high hopes that they could similarly benefit typeface design as well. But it doesn't look like there are any current practical applications.
#26 - 2023-04-05, 16:11
Raymond Larabie
Typodermic Fonts Inc.

Yes. I was quite optimistic about these new tools, as they've been truly transformative in other aspects of my work. I had high hopes that they could similarly benefit typeface design as well. But it doesn't look like there are any current practical applications.
Hi,

I have always known this, in fact since I acquired a unit of value in AI during my master degre. I have acquired the conviction that the AI is not intelligent, in the first sense of the term. It can only be a simplified reproduction of what already exists.

Thanks
A.philippe
#27 - 2023-04-06, 02:40

Practically speaking, kerning needs work.

There should be kerning subgroups (underneath the classes).

In the future, FontLab should be able to intelligently convert the pairs of that subgroup into class or exception pairs, as needed by your specific kerning.

For example, for the vyw class, there's a subclass for v, subclass for y, and subclass for w.

Also you can make a subclass for òàù and another for öüë, etc.

Then it combines that information, to make the exported kerning pairs.
#28 - 2023-04-08, 14:07
« Last Edit: 2023-04-08, 14:10 by Dave Lawrence »

There are already other better, intelligent ways of doing spacing.

For example, in sheet music Sibelius has had the "Magnetic Layout" that automatically moves things around. It has had this for more than 11 years.

Check this out:
https://i.imgur.com/4eesrtD.gif
GPT4 integration



It works like this:

You set the numbers for each sheet music element and tell it how much it should move up down or left. (A more intelligent version of this would also move diagonal!)

But these numbers are already set for you, and you only need to tweak a little.

https://i.imgur.com/Z8NNgOc.jpg
GPT4 integration


And you can make an exception for any element and move it manually to the correct space on the page.

So what would this look like in fonts? Maybe you specify the area between the H's, and then the area between the AV? or maybe you could specify the distance of the nearest points between two letters?



To be more accessible to some people, and to be faster and easier to use, font editors need to get away from their historical roots, such as kerns.

The font editors of the future will allow you to just set a minimum or average distance. (Or have a list of presets, like Sibelius!)

Then the program can interpret that as sidebearings or kerning or whatever, upon export.

And then people can share presets like "tight sans spacing", or "loose, body serif".
#29 - 2023-04-08, 14:34
« Last Edit: 2023-04-08, 15:12 by Dave Lawrence »

Then after font editors do this,

Something like this might make its way into font specs. :D

---------

Addition: Does anyone know the algorithm for how Adobe inDesign handles right and left "optical alignment"? Also, "optical" kerning?
#30 - 2023-04-08, 14:39
« Last Edit: 2023-04-08, 20:50 by Dave Lawrence »

Vice-president, FontLab

I'm not sure that any "universal" model (other than advanced AGI, but then we won't bother much with fonts) can do something good, except to offer some graphic ideas in the style of MJ. This means that before we start training any specialized AI model for fonts (and expect any human-quality results), we need to better understand what a "digital font" is.

What we have now:

1) we have an idea of the font, we know its visual and semantic structure ->
2) we convert it into a "digital font", which is a stupid set of pixels or some kind of polysplines ->
3) when it is rendered and we read it, our brain recreates the original "idea" of the font.

Any training applied at level 2 will only work if it is "mapped" to level 1, which will require a very formal way of defining it.

But if we can define "font" as level 1 data, then we may not need to store level 2 data at all by displaying it directly.
#31 - 2023-04-09, 06:09
Regards,
Yuri Yarmola

I wonder if GPT, in its current state, can help you localize Fontlab internationally. If you do that (like most web browsers do — where the language of the interface elements can be whatever the user wants, or even deduced directly from the operating system) you would be expanding your user base in a major way. Software capable of interfacing with its potential users in their own language is pretty much a slam dunk over its competitors.
#32 - 2023-04-09, 09:10
« Last Edit: 2023-04-09, 09:12 by Patrick Griffin »

Vice-president, FontLab
I imagine that our use of texts in the font development process can only benefit from using AI systems. Why not ask Char-GPT to come up with language-specific lists of kerning pairs, ordered by frequency in existing typeset texts? — If the results delivered by AI systems are reasonably reliable in this area, I think it would be highly useful to have an AI system for generating example texts inside FontLab.
Well, results are mixed. Here is my dialog with chat gpt4:

https://i.imgur.com/8H9iti7.png
GPT4 integration


I believe that algorithmic solutions may still work better to create kerning lists ))
#33 - 2023-04-09, 14:39
Regards,
Yuri Yarmola

So interesting. Thank you for having tried. :)
#34 - 2023-04-09, 15:37

Vice-president, FontLab
seriously speaking, I am very interested in anything AI, applied to fonts, at any possible level
#35 - 2023-04-09, 15:51
Regards,
Yuri Yarmola

I'm not sure that any "universal" model (other than advanced AGI, but then we won't bother much with fonts) can do something good, except to offer some graphic ideas in the style of MJ. This means that before we start training any specialized AI model for fonts (and expect any human-quality results), we need to better understand what a "digital font" is.

What we have now:

1) we have an idea of the font, we know its visual and semantic structure ->
2) we convert it into a "digital font", which is a stupid set of pixels or some kind of polysplines ->
3) when it is rendered and we read it, our brain recreates the original "idea" of the font.

Any training applied at level 2 will only work if it is "mapped" to level 1, which will require a very formal way of defining it.

But if we can define "font" as level 1 data, then we may not need to store level 2 data at all by displaying it directly.

I have no idea how AI works, but this list can go up a few levels of abstraction.

Also, this list could be expanded laterally as well.
#36 - 2023-04-11, 10:28

I have no idea how AI works, but this list can go up a few levels of abstraction.

....l.
HI,

AI is not more as any other piece of software : an algorithm applied  to data.

Artificial : yes, see above

Intelligence: no, only humans and other beings are intelligent. Machines only do what humans want them to do.

A. Philippe
#37 - 2023-04-11, 23:37

HI,

AI is not more as any other piece of software : an algorithm applied  to data.

Artificial : yes, see above

Intelligence: no, only humans and other beings are intelligent. Machines only do what humans want them to do.

A. Philippe

Hmm...I'm not sure.

What happens if you give the computer a decision making algorithm?
#38 - 2023-04-12, 00:41

Yeah, probably shouldn't go into this here more here. (I could discuss this forever!)



Getting back to FontLab, probably a good heuristic would be:

What do you as a font designer get bored with?

What do you find most tedious?

When you design a font, what feels like it takes too long?

Those would be great starting points for using AI or algorithms or whatever someone wants to call it.

#39 - 2023-04-12, 00:53

Hi,

Since 1976, nothing new :

Wirth, Niklaus (1976). Algorithms + Data Structures = Programs

https://web.archive.org/web/20110813043427/http://www.ethoberon.ethz.ch/WirthPubl/AD.pdf

Any AI is a piece of software, like FontLab

Edit :

Oh yes, something new since:
Programs + GFAM Marketing  = AI

A. Philippe
#40 - 2023-04-12, 01:11
« Last Edit: 2023-04-12, 03:06 by ALPH »


Hi,

Since 1976, nothing new :

Wirth, Niklaus (1976). Algorithms + Data Structures = Programs

https://web.archive.org/web/20110813043427/http://www.ethoberon.ethz.ch/WirthPubl/AD.pdf

Any AI is a piece of software, like FontLab

A. Philippe

This is interesting. Thank you!
#41 - 2023-04-12, 01:19

As far as FontLab.

What tasks would font designer like AI help them do?


What do you as a font designer get bored with?

What do you find most tedious?

When you design a font, what feels like it takes too long?

I'm curious about what other font designers might say.
#42 - 2023-04-12, 01:28
« Last Edit: 2023-04-12, 01:36 by Dave Lawrence »

I’ll soon publish some cool stuff that uses ChatGPT inside FontLab :) I already made a ChatGPT-based script for InDesign: https://github.com/twardoch/lorem-chatum-for-indesign
#43 - 2023-04-12, 13:14
Regards,
Adam Twardoch
Fontlab Ltd.

I like the caveats you added, Adam. :)
#44 - 2023-04-12, 14:16

Vice-president, FontLab
I was thinking about the "ideal" font editor, many years ago.

Version 1 was something which LLM do now: a dialog where you can put in some description and press "make font" button.

But then I realized that version 2 shouldn't have any user interface at all. Following what you are doing, it will already know what font you want and will do it for you when needed ))
#45 - 2023-04-12, 14:58
Regards,
Yuri Yarmola

Vice-president, FontLab
Speaking seriously, there are some tasks which I consider "easy" integrated with NN methods, such as kerning (and spacing as its compression). 

Or something like automated font categorization, or accent placement in composite glyphs (for many languages). Maybe some advanced version of FontAudit which will add NN-based analysis to check and improve outlines.

These tasks I am ready to R&D "right now".

Then there are more interesting tasks, which I don't really know how to approach, but would like to try.
#46 - 2023-04-12, 15:03
Regards,
Yuri Yarmola


Quote
These tasks I am ready to R&D "right now".


Fun and interesting as I am sure such R&D work would be, can I request prioritising missing functionality such as contextual anchors, proper filtering of mark sets in mark and mkmk lookups, Lookup panel compilation, and other things that some of us have been waiting for for a long time.
#47 - 2023-04-13, 09:21

Vice-president, FontLab

Fun and interesting as I am sure such R&D work would be, can I request prioritising missing functionality such as contextual anchors, proper filtering of mark sets in mark and mkmk lookups, Lookup panel compilation, and other things that some of us have been waiting for for a long time.
When we allocate resources to R&D we have to consider "impact" of any new feature. That's essentially a multiplication of the feature timesaving factor by number of interested users.

But we did the compilation of the Lookup panel so long ago that I don't remember why we haven't included it in one of the builds yet. Need to research it ))
#48 - 2023-04-13, 11:47
Regards,
Yuri Yarmola

Members:

0 Members and 1 Guest are viewing this topic.