Creative Fabrica - AI Font Generator

Creative Fabrica just released the AI Font Generator, is there anyone know something about this? from what i know, type design industry copyright system is very strict so how this will be work?

Comments

  • Spacing is hard.
  • Good thing I never liked those "quirky" type of fonts, because now I'm sure they will flood even more type markets.
  • While this is unrelated to the AI Font Generator by Creative Fabrica, this is the current state of business in AI font generation, as of today:


    It is not exactly "one-click" process, more like 20 clicks per font (not counting writing the actual prompt)
  • Igor Petrovic
    Igor Petrovic Posts: 318
    Without close checking, that second one looks pretty much like Optima.
  • Without close checking, that second one looks pretty much like Optima.
    That was in the prompt, "create something like Optima". 

    These samples are from very different prompts, mainly used to check current level of hallucinations and overall quality. 

    First (which looks like DIN) had DIN in prompt "create a DIN font but make it more humanistic". 

    Third is "modern geometric sans-serif", followed by "modern script", then the font that is "best for certain picture| and final font was made from some lettering (also created by AI), with prompt "create font specimen for the lettering in this picture" (or something like that).

    And this is the "blend of Comic Sans and modern geometric sans-serif" :)


  • This scares me.
  • The real issue is average users getting used to, and accepting the quality, of these. Dead internet typography all over. On the great scale we can only hope this soup of mediocre AI nonsense generates a strong enough backlash eventually, but my bet is on people’s laziness. Why think, when you can have a good enough guess? Why learn a craft, when you can fake it convincingly enough? Why spend time, when everything is available at an instant? It’s quite the promise.
  • Igor Petrovic
    Igor Petrovic Posts: 318
    edited April 23
    I am wondering if the prompt "make something similar to XY font" could technically be done without sourcing the XY typeface this way or another. I know that letterforms cannot be copyrighted, but the typeface images that AI uses are protected.

    Similar to most licensing docs prohibit making physical products based on letterforms (wooden alphabet toy, i.e.) 

    Is it a plain license infringement? Is it enough to put in our licenses that it is prohibited to use any material containing a typeface (images, etc.) for AI?
  • As a side note, please judge overall spacing quality in my font samples picture. That is relevant to the subject of this discussion ))

  • See you here again XD 
  • The mediocre stuff will improve quite fast. Remember Chatgpt two years ago? Now Claude can solve pretty hard mathematical and logical problems.

    The only obstacle is a good and labelled initial dataset, which I find not so hard to create. Actually, that was my dissertation topic. (I can share the concept if anyone is interested).

    So what now? Are the type folks going to passively look at the large players stealing their fonts and having nothing from the revenue pie?

    Maybe I am naive, but my Marxist part says that there is a chance. If type folks put forces together (whatever union) and build their own dataset, they can train their own model.


  • For now myfonts.com works as a great font dataset training machine ))
    You can "type in" any letter in any font + you have decent font family description.

    Coding the engine which will extract all training data will take a day. 

  • Simon Cozens
    Simon Cozens Posts: 766
    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.
  • This might be the reason why no such project has been published yet.
  • Alex Visi
    Alex Visi Posts: 189
    edited April 26
    Yury, are they vector drawings? Would you mind showing what “a slightly playful geometric sans-serif with a subtle contrast, semi-closed apertures, thin connections, old-style varying letter widths and an extra large x-height” would be? Your “humanist DIN” has nothing humanist about it, and the others look just like the fonts they copied, so I’m curious to see something less generic/more specific. From that post so far, it looks like a potential competition for freebies rather than for commercial foundries, but it’s interesting to see where its limits are.


    Maybe I am naive, but my Marxist part says that there is a chance. If type folks put forces together (whatever union) and build their own dataset, they can train their own model.

    If this rat race develops, it’s going to be free on google and almost free under adobe subscription, so nobody else wins this, including Yury.

  • This is what it created. Looks that it cannot be very specific yet, just reacts to some well-known words, like "geometric sans serif".



    This is bitmap. To create the font I make it 4x size with Topaz Gigapixel, then convert to font, auto-trace and auto-space in FontLab (which has "separate and trace" function for years).

    Alex Visi said: If this rat race develops, it’s going to be free on google and almost free under adobe subscription, so nobody else wins this, including Yury.

    Absolutely. Just trying to show that it's coming faster than we might expect. Expect tools like Canva to offer "custom AI-generated fonts that perfectly fit your project" within a year or less.
  • Thomas Phinney
    Thomas Phinney Posts: 2,971
    I have been saying for a long time that the leading candidate to develop “AI-generated” fonts is such a web software company providing an online design service. The most likely would be one (or more!) of those offering more low-end to mid-range tools aimed at non-pro designers, such as Zazzle or Canva.
  • Alex Visi
    Alex Visi Posts: 189
    Yury, can I also ask how large was your dataset? As far as I understand, what keeps the type industry rather safe is that the number of high quality fonts in the world is just way too small for training. Since AI needs quantity, that is 99% poorly made stuff and the results correspond. People already have access to millions of free fonts, but they still pay money for better quality.
  • It is not my dataset, just regular ChatGPT 4o. Anyone can repeat my experiment, just ask LLM for the font specimen.
  • DrayZZZ
    DrayZZZ Posts: 2
    edited 1:03AM
    I have been saying for a long time that the leading candidate to develop “AI-generated” fonts is such a web software company providing an online design service. The most likely would be one (or more!) of those offering more low-end to mid-range tools aimed at non-pro designers, such as Zazzle or Canva.
    As far as I know, Canva is already using AI-generated Chinese fonts (including the Latin part as well) in their business. 
  • Dave Crossland
    Dave Crossland Posts: 1,436
    edited 10:12AM

    Expect tools like Canva to offer "custom AI-generated fonts that perfectly fit your project" within a year or less.
    This is repeating myself as I posted this already on this forum some time ago, but I think this is really the important point in this whole discourse:

    The problem that generative A.I. presents to people interested in making fonts (I try to scope this as widely as possible and include myself ;) ) is not A.I. generated fonts.

    It is that, if the marginal cost of creating letters that work well together is zero, fonts themselves become completely obsolete and irrelevant, because document author-designers never use any font, they just generate the letters they need at the moment they need them. A file or a line at a time. 

    It's A.I generated lettering, not A.I. generated typefaces, if you will - although in a broad sense such lettering is still "mechanical writing" or whatever such definitions of typefaces versus lettering. 

    And then all this stuff about color fonts, variable fonts, wasm shaping, whatever, doesn't matter, because you just tell the authoring application you want it bolder or more multicolored in some pattern or way or whatever it would be written in your custom type brief, and it just shows it to you in a second.

    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...
  • Alex Visi
    Alex Visi Posts: 189
    edited 10:42AM

    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...
    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? 
  • Yury Yarmola
    Yury Yarmola Posts: 32
    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? 

    Same as humans do: by reading.

    Even now, there isn’t much difference whether the text is in a "readable" format (where pictures are mapped to some ID, such as Unicode) or simply a plain image without any additional data. Reading requires additional resources, but against the overall backdrop, it’s not as noticeable as it was 20 years ago.

  • Yury Yarmola
    Yury Yarmola Posts: 32
    It's A.I generated lettering, not A.I. generated typefaces, if you will - although in a broad sense such lettering is still "mechanical writing" or whatever such definitions of typefaces versus lettering. 
    I agree. And it will be an interesting cycle, from manuscripts to movable type and back again. ))
  • Alex Visi
    Alex Visi Posts: 189
    edited 1:24PM

    Same as humans do: by reading.

    Even now, there isn’t much difference whether the text is in a "readable" format (where pictures are mapped to some ID, such as Unicode) or simply a plain image without any additional data. Reading requires additional resources, but against the overall backdrop, it’s not as noticeable as it was 20 years ago.

    Sounds like an expensive and unreliable solution to a problem that doesn’t exist.