State of the art in AI image generation as we go into 2026.

Ray Larabie
Ray Larabie Posts: 1,472
edited December 19 in Font Technology
Yesterday, OpenAI released a new image generator, likely to compete with NanoBanana. Here's where we're at:

The prompt: Make a typeface specimen graphic showing the alphabet. The font is a techno typeface with capitals and lowercase.



I think the font bears some resemblance to Conthrax. The next prompt: Make a typeface specimen graphic showing the alphabet. It's the same as the previous typeface but an extra-light version.



Make a typeface specimen graphic showing the alphabet. It's the same as the typeface in the last two images but an ultra-bold version.



Make a typeface specimen graphic showing the Cyrillic alphabet. It's the same as the typeface in the last three images but the "regular" weight version.



Make a typeface specimen graphic showing a full set of currency symbols. It includes never symbols such as the riyal currency symbol. It's the same as the typeface in the last four images but the "regular" weight version.



Well, it didn't handle that last one so well, did it?  It can render some well-known fonts like Clarendon, Montserrat, and Franklin Gothic fairly accurately.

Make a typeface specimen graphic showing the alphabet in uppercase and lowercase. The background is white, and the characters are black. The font is Franklin Gothic Bold.




 I'm not thrilled about this, but I like to check in on new image generators to assess the threat. Anyway, I thought this might be interesting for anyone who hasn't been keeping up. If you want to experiment, it's just a regular ChatGPT prompt, and I think it'll work in the free version.
Tagged:

Comments

  • DanRhatigan
    DanRhatigan Posts: 43
    Ray, are these outputs just raster images? 
  • John Hudson
    John Hudson Posts: 3,565
    ‘Make a typeface specimen graphic showing the alphabet in uppercase and lowercase. The background is white, and the characters are black. The typeface is in a style of lettering you’ve never seen before.’
  • Ray Larabie
    Ray Larabie Posts: 1,472
    edited December 19
    @DanRhatigan Yes, and at 800 × 533.

    @John Hudson Here are some results from that prompt.









    Much like other image generators, it's unable to generate a hybrid of two styles like MICR/cowboy or Didone/Microgramma. Over the last few years, I haven't seen any of these tools generate a remotely interesting new typeface or typeface idea. But it's getting better at making familiar-looking typeface specimens.

  • Craig Eliason
    Craig Eliason Posts: 1,490
    Interesting that the A in the uppercase line is consistently more prominent, as if it’s a stressed initial in a piece of lettering. And that the typeface name is in letters that do not correspond to the “typeface” below. 
  • Dave Crossland
    Dave Crossland Posts: 1,515
    I guess the Photoshop plug-in from http://fontself.com will turn these into OpenType fonts, bish-bosh
  • Kent Lew
    Kent Lew Posts: 1,026
    Craig Eliason said:
    Interesting that the A in the uppercase line is consistently more prominent, as if it’s a stressed initial in a piece of lettering. And that the typeface name is in letters that do not correspond to the “typeface” below. 
    It seems the current models are more suited to creating lettering than a typeface as such. (The Fontself pathway Dave mentions notwithstanding.)


  • Craig Eliason
    Craig Eliason Posts: 1,490
    You know what though? All four of those have apt and marketable names that seem to pass <https://namecheck.fontdata.com/>! 
    Maybe after all our handwringing about AI taking our font designing jobs, it turns out it's just going to take our font naming job which we all hate anyway! :-)
  • Andreas Stötzner
    Andreas Stötzner Posts: 828
    this is turning into a pleasure …
    “create the font ‘Robomorphax Italic’ with Cyrillic, Coptic, Runic and Ethiopic!”

    “create the font ‘Cryptoblasters Cyrillic’ which design reflects the current state of Russian politics!”

    I shall stop now.
  • John Hudson
    John Hudson Posts: 3,565
    One of the things I have been testing is asking for specimens in styles of lettering that I know only exist in a relatively small number of fonts. One of the things this makes plain is just how derivative the output is, since there is less noise within which to bury the signals. The occasional ductal confusion is entertaining.

     
  • John Hudson
    John Hudson Posts: 3,565
    Interestingly, the ductal confusion gets sorted out when prompted to produce the same style in a different weight.

  • Ray Larabie
    Ray Larabie Posts: 1,472
    @John Hudson That improvement effect you're seeing is a new thing with OpenAI's current image generator. Previously, subsequent images would degrade and go off the rails quickly. But as you can see, it remembers what it's been working on and even fixes some of its own problems in subsequent variants.

    Here's a quick attempt at fancy typography.

    Generate a magazine spread for a high-end fashion magazine for the wealthy. The typography is stunning and elegant. The text is all Lorem ipsum copy.


Categories