AlphabetMagic. My first AI experiment

12467…12

Comments

  • PabloImpallari
    PabloImpallari Posts: 806
    edited January 2023
    No no.. it was about the world cup. If France had won the final match, Alphabet Magic would not exist!
  • Dave Crossland
    Dave Crossland Posts: 1,431
    😂😂😂😂
  • Die in-dryfoun
    Die in-dryfoun Posts: 27
    @PabloImpallari I may be wrong but it would not have been a healthier and more honest attitude to learn how to sketch and design original fonts to become a true type designer? I think you could have produced better alphabets than the ones you show in your examples, preserving your sanity. All this without taking into account the fact that artificial intelligence platforms are being trained on artwork that is often copyrighted.

  • Jasper de Waard
    Jasper de Waard Posts: 641
    edited January 2023
    @Die in-dryfoun If you had done a little research, you would have found that Pablo in fact designed many fonts, some of which you have probably seen many times, which he also writes about in this thread.

    Your point about copyright is an interesting one. That discussion is being held also regarding other AIs. The main ethical argument for AI, I think, is that true originality is a fantasy. Nothing is every created out of nothing. Humans reference other works in much the same way that AI does, and that has never been a problem.* Many of the examples Pablo has shown here are much more 'original' than most human-made designs.

    *Of course, there is a big grey zone here. When is a reference too strong? What if you told the AI to create a 'Serif for long-form text in the style of Gerard Unger'?
  • Die in-dryfoun
    Die in-dryfoun Posts: 27
    @Jasper de Waard It is likely that we have different tastes when it comes to type design. When I look at Pablo's designs I don't find anything that doesn't feel bland or fundamentally derivative to me. His Lobster font is nothing more than a substandard copy of Donald Young's Home Run. Its popularity is only due to the fact that it is free.
  • Ray Larabie
    Ray Larabie Posts: 1,436
    edited January 2023
    @Die in-dryfoun Pablo's work is rock solid.
  • Die in-dryfoun
    Die in-dryfoun Posts: 27
    Simon Cozens said:
    Probably the most important skill to learn in type design is not trashing other people's work.
    I'm afraid I don't share your view. I think that in type design it is much more important to do your best to design typefaces as beautiful, functional and original as possible. It's also important to be able to distinguish between mediocre designers who hate drawing to the point that they need to resort to illegal stimulants, and those who know their trade and can sketch, draw, and create new alphabets without suffering.

  • James Puckett
    James Puckett Posts: 1,998
    It occurred to me that a great set of images to train the AI on would be the contents of Rian Hughes’ Custom Lettering of the… books. 
  • Nick Shinn
    Nick Shinn Posts: 2,216
    edited January 2023
    Die etc:
    All this without taking into account the fact that artificial intelligence platforms are being trained on artwork that is often copyrighted.

    If the AI is training on scans of typography, then it is no different than type designers scanning old printed foundry specimen books, by eye or with a photographic device.

    However, if the AI is training on actual digital fonts, then the notion of “AI point piracy“ may have to be pre-empted in EULAs, if the foundries that publish fonts don‘t want that to happen. 
  • Nick Shinn
    Nick Shinn Posts: 2,216
    edited January 2023
    Here is a typical EULA stipulation, which would appear to already prohibit AI from “training” on actual digital font files:

    Licensed Users are prohibited from … creating derivative Fonts [i.e. Font Software] without prior written consent from Licensor or the Font’s copyright holder…

    And how about Open Source fonts?
    I’m not familiar with all the details of such font licences, but is it allowable to “train” on Google font software, and then publish the result as a commercial product?
  • Thomas Phinney
    Thomas Phinney Posts: 2,899
    I don’t know how Pablo’s stuff works as he hasn’t shared the details. But do read up on how Stable Diffusion works. You can think it is infringing or not, but something that did vectors, but used an approach like Stable Diffusion… definitely would not be copying point placement. Stable Diffusion is WAY more abstracted than that.
  • jeremy tribby
    jeremy tribby Posts: 251
    edited January 2023
    Nick Shinn said:
    And how about Open Source fonts?
    I’m not familiar with all the details of such font licences, but is it allowable to “train” on Google font software, and then publish the result as a commercial product?
    That's a very interesting question, given that Google Fonts use a copyleft license, the Open Font License (OFL). A rare few are Apache2. The OFL requires all derivative work to carry the same license. Training on Google Fonts to make another OFL font would be permissible (and I believe has been done by NaN)... but to make a proprietary font? I don't know. Matthew Butterick's lawsuit against GitHub CoPilot is essentially about the same thing.

    With proprietary fonts, I could see an argument that training on vectors would be tantamount to reverse engineering software, whereas we know that the design itself cant be copyrighted, so a visual approach like Stable Diffusion, as Thomas notes, is less risky... we'll get more clarity when the lawsuits start to settle I guess!
  • Dave Crossland
    Dave Crossland Posts: 1,431
    edited January 2023
    Eltra Corp vs Ringer (1978) suggests to me that processing bitmap typefaces with ML Transformers is not subject to copyright in the USA: neither inputs nor outputs. Just personal speculation. 

    And then making bezier outlines from the output bitmaps, would to me, suggest the copyright adheres the same as normal. 

    Buckle up. https://google-research.github.io/seanet/musiclm/examples/
  • Evie S.
    Evie S. Posts: 74
    The lawsuit against Stable Diffusion will be interesting to watch. In general, I agree with @Craig Eliason : There's so many examples of tech enthusiasts pointing out how good AI art is... and it's an anatomical nightmare. I'm sure with proper curation it would look fantastic, but in the wrong hands it's nothing more than a shiny gimmick. AI can be a great tool for artists, but I have to be convinced this will be more than a modernized TypeCooker.
  • PabloImpallari
    PabloImpallari Posts: 806
    edited January 2023
    Dave Crossland said:

    Buckle up. https://google-research.github.io/seanet/musiclm/examples/
    Wooowww!!
    I wonder what will be the sound of alphabets as songs!
    Also the other way around: songs as alphabets!
  • James Hultquist-Todd
    James Hultquist-Todd Posts: 420
    Dave Crossland said:

    Buckle up. https://google-research.github.io/seanet/musiclm/examples/
    Sound like a pile of mellotrons run through a sequencer. This makes me very sad.
  • Nick Shinn
    Nick Shinn Posts: 2,216
    Thomas:
    an approach like Stable Diffusion… definitely would not be copying point placement.
    I wasn’t referring to copying particular point placement. As Jeremy says, training on vectors would be tantamount to reverse engineering. Font EULAs say “you can open the hood and look but you can’t touch”, i.e. no data copying of font outlines, even for the purpose of running a group of them through an algorithm to create a composite outline.

    If blending two fonts to create a ’tween is not allowed, why would blending thousands be any different, even if the level of abstraction is more complex? Where to draw the line?
  • Thomas Phinney
    Thomas Phinney Posts: 2,899
    edited January 2023
    Nick, you are still suggesting “data copying of font outlines” would happen. I would argue that this is not what is happening now with Stable Diffusion, nor is it what I am suggesting as far as a font-oriented equivalent would work.

    What is happening now with Stable Diffusion may be infringing, but if so, it is not because it is copying pixels.

    “If blending two fonts to create a ’tween is not allowed, why would blending thousands be any different, even if the level of abstraction is more complex?”

    First, who says blending two fonts is not allowed? But for the sake of argument, let’s grant that. Blending thousands would be different precisely because it is thousands. One of the key deciding factors in copyright infringement under US law (and most other countries seem to have similar principles) is whether the infringing work is “substantially similar” to the specific copyrighted work it is accused of infringing. You aren’t going to be substantially similar to any one work if you have blended thousands. You are also probably not going to be much hurting the value or potential market of any one of those thousands of works.

    I won’t be shocked if new laws get created to deal with all this. That would be a reasonable response to the new realities, and copyright laws and principles that never imagined the situation we are now facing.

    I am not a lawyer, but the few lawyers I have seen weigh in on this who seemed to understand the tech, and were not working for one side, had no expectation for which way(s) the courts will decide, based on existing copyright law. So, I would not be even a tiny bit surprised if we got conflicting court decisions, whether between different lower courts in the USA, or between the USA and other countries.

    It could be a wild ride! We should buckle up!
  • Nick Shinn
    Nick Shinn Posts: 2,216
    Thomas, my understanding of “data copying” is that glyph descriptions are extracted from a font file in electronic digital form. Whatever happens to it next is of no consequence, in respect to a EULA which is a contract that says you’re not allowed to do that.

    However, if a third party not involved in the contract were to extract fonts from web sites, and the fonts are not software protected by encryption from being opened, then I assume EULAs are irrelevant and general copyright laws would apply.

  • Thomas Phinney
    Thomas Phinney Posts: 2,899
    OK, so it is still not at all clear to me that the kind of process I have in mind (analogous to Stable Diffusion, but for vector fonts instead of bitmapped images) matches your description of data copying. I imagine that could depend very much on the details of the particular EULA, and quite possibly on the kind and degree of abstraction from the algorithm.

    And of course at no point did I or anybody else in the discussion mention EULAs; I assume that this will be web based and/or programmed by people not party to any relevant EULAs, so that part won’t matter. At least, not if the people involved are at all smart about it.

    Another consideration is, I expect releasing fonts generated this way might very well trigger threats of legal action and likely actual legal action, from at least some of the major players, if they believed their fonts were part of the training corpus for the font-generating engine.

    Anyhow, apologies to @PabloImpallari for hijacking his thread! I don’t need to go on further.  :pensive:
  • PabloImpallari
    PabloImpallari Posts: 806
    edited January 2023
    Thomas, Nick & all: fell free to discuss here whatever you want related to AI.. I have no problem with it.

    Quick update on the "Testing Page" experiment:
    Training the network for "top quality" contours as I would like it too, would have taken 40 hours.. to much for an experiment. I reduced it to a more acceptable time of 4 hours. It was about to finish training (96%) when I got kicked out of Collab since Im on the free plan and they give preferences to paid users.

    I will try again tonight when everyone is sleeping and there is more time available for free users
  • PabloImpallari
    PabloImpallari Posts: 806
    edited January 2023
    Nick Shinn said:
     Where to draw the line?
    Great Question. 

    When I started making fonts 13 years ago, most fonts where single style, or RIBI. 
    The Univers table was a rare thing... nowadays multiaxis fonts are the standard, released everyday.

    That was possible thanks to the progress in tools like Fontlab Multiple Master, Superpolator, RMX Tools, GlypphsApp, etc. AFAIK, nobody asked "Where to draw the line" at that tine.. maybe we should have asked... I don't know...maybe it was a good thing we didn't asked... who knows? Even 1st timers are doing families from day 1.. Is it a good or bad thing? I don't know.. maybe is good.

    Now with AI we are able do more than axis expansions. For example for the mayans and the monster experiment I played with dingbats. For the "seashore inspired number2" I played with this one

    Maybe we should ask, as you did, where to draw the line... or maybe we can keep using the same line as today, releasing new fonts that wont hurt the value of our colleagues fonts as much as possible.... However that has happened many many times in the industry since the beginning of metal type, without computers or AI involved.

    My initial guess is that AI will mostly expand the range of display alphabets, as Multiple Masters and interpolation did with multi-axis families. Text fonts are already very similar to each other, hence really hurting the value to each other.. even if drawn by hand or node dragging whit no AI involved.

    Also, I don't see many font descriptions telling "we have used interpolation and rmx harmonizer to develop our families" or any other tool, whatever it may be... for example: iKern is used a lot, but only a few are telling it while others keep it private. AI is just another tool. There is nothing in any particular released font that can tell if you have used rmx tools, or fonttools or whatever any other tool, or not... I guess the same will apply to AI fonts: they will be AI fonts as long as you say so in the font description... or not
  • Thomas Phinney
    Thomas Phinney Posts: 2,899
    Evie S. said:
    … In general, I agree with @Craig Eliason : There's so many examples of tech enthusiasts pointing out how good AI art is... and it's an anatomical nightmare. I'm sure with proper curation it would look fantastic, but in the wrong hands it's nothing more than a shiny gimmick. AI can be a great tool for artists, but I have to be convinced this will be more than a modernized TypeCooker.
    I agree with everything about the present, and yet… there is every reason to assume that the tech will improve. We have seen it many, many times with software and computers doing new tasks.

    So yeah, we are at the version 1.0 stage right now and there are things about it that are laughable. But I would eat my hat if human anatomy (at least in common positions) was not mostly solved in another year or three.

    Maybe horses and dogs will still suck for a while longer after humans are solved, because they are going to be a lower priority. But even for that I think… give it time.
  • jeremy tribby
    jeremy tribby Posts: 251
    Ray Larabie said:
    If a new version of Monotype Grotesque were commissioned today, it would likely be turned into interpolated masters and most of the charm would be removed. I think the inconsistency is its best quality—it doesn't make a lick of sense, and that's what I love about it. It's possible to build an interpolated typeface with those inconsistencies, but it's unlikely that it would even happen.
    Not to get too far off topic, but Community Gothic from Frere-Jones is an example of a type family that (IMO) rather successfully achieves what you're describing. I do think you're right about the effort involved, I can only imagine that the family was particularly painful to produce
  • Ray Larabie
    Ray Larabie Posts: 1,436
    edited January 2023
    “breed of talentless, self-perceived font makers”
    I think it's worse than that. Even those won't be needed. Why would the end user bother buying or downloading a font when they could simply type in a prompt? Even if Pablo doesn't succeed with this, someone will. I can't envision a scenario where a third party is required. User requests a font, AI generates it, transaction over. It's fascinating, and it sucks.

Categories