AlphabetMagic. My first AI experiment

Options
13468911

Comments

  • Ray Larabie
    Ray Larabie Posts: 1,386
    Options

    Experiment 13: It's difficult for me to tell if the AI is copying or reaching its own conclusions. A more unconventional concept would show its capacity to harmonize odd shapes. Please program this thing with some courage and get it to generate some strange typefaces. These examples are simply too commonplace.

  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Ray Larabie said:
    Please program this thing with some courage and get it to generate some strange typefaces.
    <3 <3 <3 <3 <3 <3 <3 <3 <3 <3
  • James Puckett
    Options

    Please program this thing with some courage and get it to generate some strange typefaces. These examples are simply too commonplace.

    Agreed. I want to see this do something as wild as Simian, Blackcurrant, or Degrassi.
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options

    Experiment 13: It's difficult for me to tell if the AI is copying or reaching its own conclusions. 

    It's reaching his own conclusions!  :#

    Given enough number of iterations all the alphabets end up looking like Gerard Unger's alphabets, even if Gerard's alphabets are not included in the training.. and by that I mean that their shapes get reduced to primitives.

    Does anyone have Gerard's book "While you are reading" at hand? My copy is inside a storage deposit and I cant get it right now. To explain what I mean without misunderstandings I will need if any of you can take photos of the the translucent pages inside Gerard's books: The ones with basic letter parts that sit in front of each others creating the full letters. Because that''s exactly what I'm getting out from Alphabet Magic when I increase the sardines swim time long enough.

    I have read his book 4 times and was ashamed that I was unable to fully understand all of it..... now by experimenting with AM, I finally do. It feels like Gerard has had superhuman brain powers!

    I would like to post some images of letters produced by AlphabetMagic that blends together in the same way that the translucent pages do in Gerard's Book, or as in Gerard's Decoder font. But I'm afraid that people wont believe it unless they have seen the book.

    Also, some other letterforms that AM generates looks similar to alternatives inside Alberata, you can see some of them in the result from experiment #1. The /E blending into a /C and so on...

    If you keep increasing the number of iterations beyond Gerard level, they get reduced to basic triangles, circles and squares: For example The Letter P blends into the Letter D, then also the D gets transformed into letter O and so on... the /h ascender start to shorten into /n shape, and so on... the /b blends into the /o... etc... You get the idea.

    In Experiment 13 artwork... iterations are at the limit.. the /p was already starting to blend into a /b..   If you open experiment #13 in GlyphsApp and start to play with the shapes on top of each other you will start noticing some of the Unger "while you are reading" vibes inside it, even when not any Unger alphabet was used in the experiment.
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Experiment 13b is up:

    Before going crazy as Ray and James suggested -I cant wait!- a quick variation of #13 with a relaxed feeling and more influence of the french alphabet so its much more easier to guess this time. Same process as #13 but using a new auto-tracer: We have great contours now!





    John Butler guessed Jenson: 7.7% of it. What about the other 92.3%?

    Let's throw a challenge:
    If someone guess the 100% of the recipe, the winners get to decide the future of AlphabetMagic. I can gift a copy of AlphabetMagic to the winner, or make it open source, or erase it from my computer, or whatever the winner wants. Contest ends Monday morning.

    Ready, set, go!
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Ray, James <3<3 .... You wont fucking believe it!
    Give me 1 hour or 2 to post the results.... I don't even know where to start!!! ...
    or ping me into skype, my username is pabloimpallari or pablo impallari or something like that
  • John Hudson
    John Hudson Posts: 3,017
    Options
    Looks like 13b contains a lot of the same DNA as Robert Slimbach’s Brioso.

  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Hi John. Yes, it is... I made it pretty obvious!  o:)

    But there is a catch... or a concatenation of catches:
    • Both 13 and 13b shares the same recipe. Brioso was impossible to detect in 13(a)
    • A tinny detail for the obsessives like me among us: Its not "exactly" Brioso.. its KazurakiSP2N-Light. The weight is slightly different... but of course is brioso too.
    • I haven't increased the importance of KazurakiSP2N into the mix.. you only noticed brioso now because what I increased is the importance of the -so far mysterious- french alphabet, and Brioso is closely related to it. I wont go as far as saying that Brioso is a copy because is not, but is sort of a "redraw by hand" heavily inspired in the french alphabet, that since no one is gonna guess it because is pretty obscure, I will name it: The french alphabet is: ..... drumroll.... Boltana's Rabelais!... French calligraphy as it best!
    • And so, by increasing Boltanas influence over the mix, Brioso showed itself.
    Ok.. so far:
    1. Jenson 7,7%,
    2. Brioso 7,7%
    3. Bolatana 7,7%
    Thats only 23,1%. The other 76,9% remains to be identified... and in all honestly I thinks is pretty much impossible. I can extend the challenge time a whole week, if someone really want to try his hand at it.

    I have made this little identification game as a sort of meta-experiment. So we all get the experience of what will most likely to start happening sooner or later, once AI goes mainstream into type design. It's a learning experience for all of us.

    There are 13 alphabets involved in experiment 13 (the same for both a and b) and is already quite difficult... can you image mixing 100 alphabets, as Thomas pointed? Its and entirely new game.

    As Ray said: it's fascinating and sucks!
    Or dear friend, fondue!, was worried about the untalented using AI... let me tell you something...is the same if you give them AI or if you give them a good old H2 pencil... the really talented ones are more dangerous!! Isn't it?

    By the way John, I wanted to ask you:
    What's you most honest opinion regarding experiment 13(a). I selected the alphabets to be included in the recipe by thinking of you, but also trying to please me too. Trying to pleas both of us, so to speak. Of course you know tit is an AM generated alphabets.. but try to imagine: If someone else, not me, but some else.. shows you experiment 13 lying to you as if it was drawn by herself. What would you have said in a event like that? In all honesty?

    Please program this thing with some courage and get it to generate some strange typefaces.

    Experiment 14 is up!
    Warning: I went really cockoo bananas on this one... go see for yourself and have fun!

  • John Hudson
    John Hudson Posts: 3,017
    Options
    By the way John, I wanted to ask you: What's you most honest opinion regarding experiment 13(a). I selected the alphabets to be included in the recipe by thinking of you, but also trying to please me too. Trying to pleas both of us, so to speak.
    I’m not sure what you mean by ‘thinking of me’. Presumably some idea of what you think I would like? But that’s not at all the point of my critique, which is about process. What the AI is doing is not type design because it is guided only be external references and not by internal reflexion, so the choice of alphabets as input is sort of irrelevant to what I find problematic. The results seem to me boring—and I don’t just mean the ones that you have tried to make boring—because there is nothing in them that exhibits the processes that make typefaces interesting. Now, there are plenty of typefaces produced by humans that are uninteresting and just not very good for the same reason, so it’s not that I am just being critical of AI. :#
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Ok... this is what I'm understanding from reading you:

    1. You don't like the process of using AI because you think there is not internal reflection, and is not type design.
    2. You also don't like the end results because you don't see nothing inside the resulting alphabet that shows a internal reflection process too.
    3. You are also being critical about me because you think I'm not doing type design, I'm doing something else.. whatever that may be.
    4. You also don't like the result from any experiment for the 3 previous reasons.

    Can it be summarized like that? Are we both in the same page?
    If so, I will like to argue that, as long as you are interested in doing so and open to new and different ideas.

    I can't explain it...  but somehow I have a gut feeling that I can change your mind
  • James Puckett
    Options
    I’m impressed by the autotracing of 13b. The techology has come a long way.

    14 is interesting. But it seems like most of the wild creative shapes are illegible; some of the glyphs look like an alphabet for communicating with angels that you’d find in a witchcraft book at an occult shop.
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    I warned it.. I went really crazy with it since Ray asked for "strange" typefaces.
    I think I pushed it too much...
  • John Butler
    Options
    Brioso jumped right out at me as well. Perhaps a bit of Nicholas Cochin in there too.
  • John Butler
    Options
    and Vendetta for the lowercase y
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Not John, this time you didn't nailed it. Is not that easy as it seems  B)

    I hope you don't mind about the the clicks on the agree/disagree buttons.... I will be using those as "Yes"/"No" for the type identification challenge. I think I will also come back later, once the challenge finishes, to remove those clicks, so they don't show up in your profile.
  • John Butler
    John Butler Posts: 259
    edited February 2023
    Options
    Doh. I had a perfect record of zero disagrees and strikes until just now. :-( I withdraw my guesses.

    Edit: never mind, didn’t get your response the first time.
  • PabloImpallari
    Options
    hahahaha keep guessing as long as you want... if you don't get an Agree click, its a miss :)
  • John Hudson
    John Hudson Posts: 3,017
    edited February 2023
    Options
    1. You don't like the process of using AI because you think there is not internal reflection, and is not type design.

    It isn’t a question of liking or not liking the process. I just don’t see it producing interesting results because it doesn’t go through the steps that I think would be necessary to produce interesting results. All it is doing is producing more-or-less clumsy hybrids of other designs. That isn’t interesting, despite the technical cleverness on display.

    2. You also don't like the end results because you don't see nothing inside the resulting alphabet that shows a internal reflection process too.
    I am not impressed with the results, and yes this is related to my critique of the process in that I think the process is missing the steps that would be needed to make the results better. Any of your experiments could be the start of something interesting and worthwhile, but their reliance on external models also means that they are disadvantaged in that respect compared to, say, a beginner student attempt at type design, which might be similarly clumsy without also being derivative. So even if they are improved iteratively and using a model of internal reflexion, they are still likely to show signs of their derivation.

    3. You are also being critical about me because you think I'm not doing type design, I'm doing something else.. whatever that may be.
    No. Let me be very clear about this: I am not being critical of you, and I am not critical of what you are doing on the grounds that it ‘isn’t type design’. I would say, perhaps, that it isn’t yet type design, but still I would not be critical of it on those grounds. I am only critical of the results so far, as I would hope you are too, and trying to understand why they might not get significantly better within the current process. I am trying to be helpful, not critical, because it is obvious that this technology is going to play some kind of role in making things that used to rely on human craft, and my concern is that the current enthusiasm is properly tempered with critical analysis of both process and outcomes. 
    4. You also don't like the result from any experiment for the 3 previous reasons.
    Again, it isn’t about liking. Let me acknowledge the technical cleverness on display, and your own grasp of the technology and ability to work with it, as exhibited in the experiments. Compared to some early attempts at AI lettering, these seem a remarkable improvement. But I am not the sort of person who is inclined to applaud and say ‘Amazing!’ because something is technically impressive.

    You have referred to these as ‘experiments’, and I am treating them as such. Part of the process of experimentation is observation and conclusion, and I think this is what I—and some of the other people involved in this thread—are engaged in, when we ask questions about the process, about how it works, suggest additional experiments, and suggest what might be needed to produce better, more interesting results.
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    That's the point... you don't see it... just yet.. but you will be able to see it once you are able to experiment with it.

    I do agree with you that I'm not doing Type Design... I am doing something else, is not calligraphy, is not lettering, is not nothing... is something entirely now and we don't have a name for it. I disagree that there is not a internal reflection, I think that there is, only that is a different kind of internal reflection.

    I had problems understanding your previous comments, maybe, because I was expecting something more like a standard type critic: for example something like "I don't like your result because your /m looks wide" or something more basic like that... something more tangible to me... but of course you don't have any obligation to do it the way I would have expected... you have your own style and now I'm able to finally understand your the way you are and also to appreciate it even more.

    I also had all this internal struggle inside mi.... is this AI shit good, it is a disaster...what should I do with it? I have asked many times and nobody have responded that.. maybe all I want is some that tells me: "erease it from your computer once and for all"  because I know that once this shit gets out to the public...  in hands of people that only worries about money that will begin to sell font duplicates the very first days... or in the hands on monotype... I do know inside me that this will be a catastrophe... that's why I haven't shared it yet. Also, I havent used it to release new fonts.. even when that would have been very easy for me to do... I do know that it fells like its not the right thing to do... it fells like cheating.

    But also I feel fascinated by the power of it... The words by Ray keeps resonating in. my head: its fantastic and it sucks! So please John... your tell me: What should I do with this shit?

    As for me... I will be forever happy only by playing with it as in these little experiments and just looking at the crazy alphabets artwork that gets created... I don't even have the need to autotrace and to put it inside a font... I'm just totally happy looking at the pngs, in the same way that I am happy when I look at old specimen books... thats why I am sharing the experiment.... for me it feels like sharing an old specimen.... If you read again my initial post.. I talked about the Dan Solo catalogs.

    I have only turning these alphabets into fonts.. becaase all people here wanted to test it on the testing page.

    Don't you have curiosity about what will be the lovechild of... I dont know... a blackletter and helvetica... or silly funny little things like that? For me that stupid little things are super interesting!
  • John Hudson
    John Hudson Posts: 3,017
    Options
    What should I do with this shit?
    I think there is probably a usefulness in rapidly exploring parametic suggestion fields for design ideas. I can imagine using it that way: as a kind of filtering system to quickly figure out what directions might lead to something interesting and useful, and which look like dead ends.

    For some people, AI might well be an exploitable substitute for imagination, but it could also be a tool for training the imagination. I suspect that if you do enough of this kind of experiment, you will become pretty good at anticipating the results, because at the same time you are ‘training’ the AI, it is training you in how it works and what it is going to do with a given input. So, the more you use it, the more easily you will be able to imagine the results before they happen.
  • jeremy tribby
    jeremy tribby Posts: 225
    edited February 2023
    Options
    One question that has been on my mind throughout this thread: why wasn't Prototypo more popular? Not enough parameters? Not enough utility? Other?

  • James Puckett
    Options
    I never messed with Prototypo because I’m a glutton for punishment and I’d rather grind it all out myself. Also I rarely have paying clients so I don’t need to cut corners and just get a type family finished in two weeks.
  • Ray Larabie
    Ray Larabie Posts: 1,386
    edited February 2023
    Options

    Experiment 14’s results are intriguing, but far from practical. This is comparable to the roadblock I encountered while exploring with Dall-E 2 and Midjourney. When I was able to create traditional designs, the outcomes were legible and consistent. The results of its unorthodox designs were illegible and inconsistent. I believe that crossing that line is the key to make this tool useful rather than amusing.

    Experiment 14 fails even as an idea generator, since there are so few examples of several legible characters that appear to be from the same typeface. Humans are exceptional at determining legibility. Designers in display type frequently straddle the line. My Zelega Zenega is an extreme example. The legibility is pushed to its limit, but it remains equally unintelligible throughout the font. The consistency of legibility is one thing that distinguishes a useful font from a useless one. It’s the clearest point of failure that I can see.

  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    I agree with all that you have wrote guys!
    We are pretty much arriving all together to the same conclusions. Thats good!

    is producing more-or-less clumsy hybrids of other designs
    Still, let me try to clarify this "more or less clumsy hybrid" concept.

    I believe It is doing much much mora than that..... its more like sort of "a 100% perfect distillation" rather than a "hybrid"..... but there is also a difference.

    Its also not a distillation... because when you distill whisky or vodka or whatever... the distillation process filters out all the imperfections.... as humans do when they do type design. I think thats the process that John is talking about.

    The AI in the other hand is doing a perfect mix of both "the virtues" and also "the defects" of the alphabets included into the mix. Thats the difference to what humans do.

    Bt humans also do mistakes... so maybe what the AI is still doing the same as humans do... I don't know... is perplexing to me....

    I guess you are still not able to see the same little things that I'am seeing, because you don't know the recipe than I'm using for each experiment.

    Maybe.... in order for all of us to be in the same page... to have a better understanding of each experiment... and to be able to evaluate each experiment whit "all the data" we all need to know both things: The recipe of the alphabet that went into the mix AND the results... all together.... as opposite to what I have being doing until know... when I'm only sharing the experiment results but not the experiment recipe.

    Would you guys be interested in doing such experiment, knowing "all the data"?
    I guess that once we do that... we all are gonna be "in the same page" and able to evaluate the results in a different light. What do you think guys?
  • John Hudson
    John Hudson Posts: 3,017
    edited February 2023
    Options
    The AI in the other hand is doing a perfect mix of both "the virtues" and also "the defects" of the alphabets included into the mix.
    That doesn’t seem to me to be the case, based on looking at your results. Or, at least, not only the case. All the experimentes exhibit what seems to me a general level of clumsiness, which I think is coming from the cross-breeding of elements of the different input alphabets in the mix without any way to refine the internal consistency of the output. The relative proportions are less good, I suspect, than that within any of the input alphabets, and the curvature and stroke modulation wander around a bit and don’t settle into a particular pattern that helps establish the character of an individual design because these are not individual designs. The trouble with combining references to different things is that all the differences between those things are a potential source of inconsistency in the combination. Yes, they can also be the source of interesting and unexpected shapes, but without the refinement of the internal consistency the results will tend to look clumsy.
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    Yes John and Ray and James and all!!!!
    I understand you 100% now!!!!

    In order to understand you better, this time I'm running and new experiment but not on the AI... this time Im running a experiment on myself. I have put AlphabetMagic to sleep, and returned to our old friend, node dragging - wich I no longer hate <3 .

    Experiment 15: AlphabertMagic vs Myself, goes like this:
    I selected one of the artworks that I quite liked from experiment #1 (512px image) and started to work on it.. old school style.. and you are totally right... now I get it.

    This was the initial artwork import after auto-tracing (and proper spacing):


    This is what it looks like at this time, after 76 node-dragging iterations:


    Everything that you just wrote... I was experiencing it at the same time, all day long today.. it is like our minds where synchronized!!

    I have started working on this one on Saturday 11am, non stop until now (Sunday 2:30am).
    So far in 15 hours achieved 76 iterations: And all I have done is only overall spacing and letters: "a b d e g h i l m n o p q r s".... progress is a bit slow: 15 letters in 15 hours = 1hour x glyph

    By the way... I'm using your "nihilim" method.

    If you want to see all my little iterations, my workflow process and my design decisions, all the 70 .off files for each iteration are inside the "My node dragging Iterations" folder.

    I'm going to try to sleep now, and continue tomorrow.
    Thanks for all your advice guys... all of you! And let me know what you think!
  • John Hudson
    John Hudson Posts: 3,017
    Options
    Nice. Seeing the two side-by-side really exposes both the limitations of the AI output and also its usefulness as a kind of rough sketch stage. The texture of the iteratively revised design is so much better, and the decisions you made about how to flare the extenders, how to balance the counters and apertures, etc. make the design more interesting.

    [Your lowercase r seems not quite right: I understand you want to keep the texture tight, but the arm seems too tentative and almost clipped looking. It needs a more definite shape that harmonises with the softer terminals in other letters.]
  • PabloImpallari
    PabloImpallari Posts: 783
    edited February 2023
    Options
    ... the limitations of the AI output ...
    Please, allow me to replace it with:
    ...The limitations of the current state of the arts of AlphabetMagic version 1.0...

    The voices inside my head are telling me that AMv2 can also be trained to perform all the type designer's typical node-dragging iterations too. I think I will be able to train AMv2 to perform a "Make it interesting!" experiment: Technically speaking is not too much different to the "Make it boring" experiment.

    Everything that can be expressed as a "decision making process" can be trained into the AI, and the AI will eventually get better than us at doing it, and Type Design is all about decision making, so...  shivers :# !!!!

    Also.. the more node-dragging iterations and the more "interesting" the design gets: the more it looks like a computer has made it....  On the other hand, the initial imported artwork looks more "human made" in the sense that it looks more like a woodcut alphabet.... despite being "clumsy or clubbed together".....    Perplexing :'(  !!!
  • John Hudson
    John Hudson Posts: 3,017
    Options
    the more node-dragin iterations and the more "interesting" the design gets: the more it looks like a computer has made it....  On the other hand, the initial Artwork produced by AMv1 looks more "human made" in the sense that it looks more like a woodcut alphabet
    I don’t think a woodcut looks more ‘human made’ than something made by a human on a computer. They’re just different media with their own characteristics. There’s certain slickness to rasterised beziér type, antialiased at high resolution, that is hard to overcome and affects a similar ‘digital feel’, and I think type designers have gravitated towards embracing that feel in the methods they use to define outlines, e.g. using harmonised node placement and balanced control handles. I suppose one could say that a lot of recent type design has been ‘digitally inflected’, and this is what you are responding to when you think the iterative results looke more ‘computer made’.