AlphabetMagic. My first AI experiment

1…67891012»

Comments

  • John Hudson
    John Hudson Posts: 3,204
    edited August 2
    One can produce some quite lovely results if the target style leans on a kind of rough informality, e.g.


    Regretably, the model generates each letter individually, so there is no overall consistency across the alphabet.


    I do like the Dutch IJ that this produced instead of J, but that is the sort of weird error that I have also seen result in an occasional backwards L.

    And as ever increasingly with AI, it is horrific to consider the amount of energy being expended on this entertainment.
  • PabloImpallari
    PabloImpallari Posts: 806
    edited August 6
    Flux, a new image generator AI  (similar to Midjourney or StableDifussion) has been released, and its open source. People comments around the web are saying thats it's better at generating text inside images, but I haven't tested it.

    For the prompt "Create a new typeface, show letter font A to Z in uppercase, in a romantic style" I got this result:

    And prompt: Write "Computer calligraphy" in the style of john steven



    Another example


  • Dave Crossland
    Dave Crossland Posts: 1,430
    "If you have a well-equipped laptop you can download and run Flux.01 locally."

    – https://www.tomsguide.com/ai/ai-image-video/forget-midjourney-flux-is-the-new-king-of-ai-image-generation-and-heres-how-to-get-access
  • Dave Crossland
    Dave Crossland Posts: 1,430
    edited August 28
    Flux demos of very good handwriting and lettering style transfer

    https://x.com/fofrAI/status/1827474269170725350?t=Z0TRZNTYYHu0YuvhBQI1cg&s=19 (now deleted)

    https://x.com/jfischoff/status/1827303667453583548?t=mlVNiyMoYbnTJuJHSGj_Ug&s=19

    Media piece about it: https://arstechnica.com/information-technology/2024/08/hobbyists-discover-how-to-insert-custom-fonts-into-ai-generated-images/
  • PabloImpallari
    PabloImpallari Posts: 806
    Now that we can upload images for AI to see, I wanted to test how much typographic information it can detect from a sample image... the result is impressive.

    I uploaded the following image of Manfred Klein's Petit Script, this is the image:

    And this was my prompt:
    "What can you say me about the font used in that image? I'm a type designer and want to know everything you can say about it. Write extensively. Thanks!"


    This is GPT4 reply:

    The font in the image you’ve shared appears to be a classic serif typeface with characteristics reminiscent of historical calligraphic styles, possibly inspired by early modern humanist typefaces. Here’s an extensive breakdown of its features:

    1. Historical Influences:

    • Calligraphic Roots: The style evokes the essence of humanist minuscule, which evolved from the classical Roman alphabet but with more flowing curves. The humanist scripts of the Renaissance period (15th century) are clearly echoed in this typeface, especially in its organic, hand-crafted letterforms.
    • Venetian Influence: It resembles Venetian Old Style typefaces, which were among the first Roman type designs in the late 15th century, particularly the works of type designers like Nicolas Jenson. These fonts were known for their elegant yet sturdy construction, balanced proportions, and strong contrast between thick and thin strokes.

    2. Letterform Analysis:

    • Serifs: The serifs in the font are slightly tapered and bracketed (transition from the serif to the stroke is curved), which softens the rigidity of the letterforms. This makes the typeface feel both formal and approachable.
    • Stroke Contrast: There is noticeable contrast between the thick and thin strokes, typical of old-style typefaces. This feature emphasizes the influence of broad-nibbed pens used in calligraphy. The thick parts (vertical strokes) are heavy, and the horizontal strokes or curves are thinner, creating an elegant rhythm in the text.
    • Terminals: The terminals are often rounded or slightly flared, suggesting a calligraphic origin. These add a decorative yet refined touch, especially seen in letters like "a," "c," and "f."
    • Italic Characteristics: Even though this is a Roman (upright) type, some italic characteristics can be seen, especially in the slanted terminals of letters like "f," "j," and "y." These suggest that the designer might have drawn inspiration from cursive forms as well, blending both upright and cursive styles.

    3. Specific Letter Features:

    • Uppercase:
      • The uppercase letters display sharp serifs and clear, bold forms. The 'A' has a wide stance, the 'E' shows a strong horizontal emphasis, and the 'G' has an open counter, giving it a classical Roman inscriptional feel.
      • Some letters, such as "Q," have a pronounced tail, a characteristic common in more decorative and classical serif fonts. The tail of the 'Q' is flowing, almost resembling handwritten styles.
    • Lowercase:
      • The lowercase has a distinctive balance between roundness and sharpness. The 'a' is double-story, typical of humanist and old-style fonts. Letters like 'g' and 'y' show descenders that are long and graceful, again hinting at a calligraphic influence.
      • The letter 'f' has a particularly elegant overhanging crossbar, which interacts well with its neighboring letters, showing a close attention to kerning and overall text cohesion.
    • Numbers: The numbers (if present in the full typeface) would likely follow old-style proportions (non-lining), where some numbers (like 3, 5, 7, and 9) have descenders and ascenders, mimicking the flow of lowercase letters.

    4. Application and Usability:

    • Text Readability: The contrast between strokes, bracketed serifs, and balanced proportions make this typeface highly readable in both large and small sizes. It would work excellently in body text for books or long-form reading materials.
    • Display Use: The refined elegance and classic feel also make it suitable for headings, invitations, certificates, or branding that requires a touch of tradition and sophistication.
    • Versatility: While this font is reminiscent of historic typefaces, its timeless quality would suit modern applications that demand a classical touch. It’s not too ornate, so it can still be considered versatile for contemporary graphic design work.

    5. Typeface Category and Style:

    • Old Style Serif: This font belongs to the Old Style Serif category, typically associated with the humanist and Garalde styles. These fonts were developed based on the shapes of letters written with a broad-nib pen, resulting in diagonal stress and gently curved, bracketed serifs.
    • Notable Comparisons:
      • It could be compared to Garamond (especially the early versions of Claude Garamond’s designs) and Bembo in terms of its organic flow and old-style characteristics. The font in your image also bears similarities to Jenson and Goudy Old Style, which are similarly rooted in Renaissance letterforms.

    6. Modern Usage:

    • The aesthetic of this font makes it suitable for branding that aims to evoke a sense of history, academia, or luxury. This includes applications in book publishing, high-end editorial design, or even luxury goods packaging where a classical, dignified look is desired.
    • The legibility at various sizes ensures that it works well across digital and print mediums. It can be applied in user interfaces, where the focus is on elegance and readability, or in formal print materials such as invitations, brochures, or corporate branding.

    7. Potential Font Name:

    • While I cannot definitively name the exact font without access to a font identification database, fonts with these characteristics might include classic names like Adobe Jenson, Centaur, Sabon, or ITC Berkeley Oldstyle. If it is an original design or custom typeface you’re working on, it could fit into this venerable tradition of Renaissance-inspired typography.

    8. Type Design Considerations:

    • As a type designer, you may want to consider the balance between stroke thickness, especially how it affects readability at smaller sizes. The thin parts should retain their elegance without becoming too faint.
    • If you're working on a typeface with this aesthetic, attention to the junctions where strokes meet (especially in letters like 'n,' 'm,' and 'h') will be key to maintaining a smooth flow in continuous text.
    • Also, maintaining a harmonious relationship between uppercase and lowercase letters is important for a cohesive overall look. As the font you shared already demonstrates, ensuring that serifs and terminals complement each other across both cases enhances visual consistency.

    In conclusion, this typeface is a beautiful example of an Old Style serif with calligraphic roots, perfect for projects that aim to capture a sense of history and elegance. If you're creating or refining a typeface in this style, paying attention to the rhythm, balance, and subtle calligraphic touches will be key to its success in both modern and classical applications.

  • PabloImpallari
    PabloImpallari Posts: 806
    Same image and prompt, this time on Gemini:

    Analyzing the Font in the Image

    Based on the image you provided, the font appears to be a variation of Apple Chancery.

    Apple Chancery is a cursive script typeface designed by Steve Jobs and Joanna Hoffman for Apple Computer in the early 1980s. It was initially intended for use in the Macintosh operating system, but it quickly became popular for its elegant and flowing style.

    Key characteristics of Apple Chancery that are visible in the image:

    • Cursive style: The letters are connected in a fluid, handwritten manner.
    • Italic slant: The entire alphabet is slanted to the right.
    • Looping ascenders and descenders: The letters with tall ascenders (e.g., 'b', 'd', 'f', 'h', 'k', 'l', 't') and long descenders (e.g., 'g', 'j', 'p', 'q', 'y') have distinctive loops.
    • Varying stroke width: The thickness of the strokes varies throughout the letters, creating a sense of movement and dynamism.
    • Distinctive letterforms: Certain letters have unique shapes, such as the 'a' with its double loop and the 'f' with its elongated descender.

    Possible variations:

    While the overall characteristics match Apple Chancery, there might be slight variations due to:

    • Font weight: The font could be regular, bold, or italic.
    • Optical size: The font might be designed for a specific size range (e.g., small, medium, large).
    • Customizations: The font might have been modified or customized to fit specific design needs.

    Additional considerations:

    • Historical significance: Apple Chancery has a significant place in the history of typography, particularly in the context of personal computing.
    • Popularity and usage: It has been widely used in various design applications, from branding and advertising to packaging and web design.
    • Digital availability: Apple Chancery is likely available as a digital font file, allowing designers to use it in their projects.

    To provide a more accurate identification and analysis, I would need additional information, such as:

    • The specific context of the image (e.g., where it was found, the subject matter).
    • Any accompanying text or metadata that might provide clues about the font.
    • A higher-resolution version of the image for a more detailed examination.

    With more information, I could potentially determine if the font is indeed Apple Chancery or a close variation and provide a more in-depth analysis of its specific characteristics.



  • PabloImpallari
    PabloImpallari Posts: 806
    edited September 10
    Craig Eliason said:
    You posted that to show off results you regarded as "impressive," but to me it's a good example of the bullshitting I've come to expect from ChatGPT. The sample is nothing like Nicholas Jenson's types, and is not a "Roman (upright) type." /G having an open counter is an odd observation (maybe an open aperture?). Asserting a "classical Roman inscriptional feel" doesn't seem right to me at all. Disagree that it would work "excellently in body text for books."
    This reads like "type design talk" but strikes me as disconnected from the actual visual prompt.
    I think you get me wrong.. or maybe I didn't explain myself correctly.

    For me they are impressive compared to just a few month ago, when all we were able to do is to feed Dreamboth with a little image (of a limited amount of pixels), and all we got in return was something ultra limited like "The image is a charset from a font" or something quite basic like that.

    The results we get now are far from being perfect, we all can agree on that. But the t    he progress and the speed at which this things are advancing is impressive to me. When I look at the results of all this AI experiments, I don't look at them as "finished products" but as "work in progress" instead.
  • Nick Shinn
    Nick Shinn Posts: 2,210
    edited September 10
    Presumably, this site will be renamed FontPrompters.
  • Stuart Sandler
    Stuart Sandler Posts: 362
    edited September 14
    @Ray Larabie you're referring to what's known as AI hallucinations and the closest antidote to date on the info side is Perplexity.ai which cites all sources of information however it does attempt to fill in the blanks when pressed for details makes inaccurate guesses. That said, Adobe is definitely making progress on the design side and it will remain to be seen how they address non Adobe Fonts - https://en.softonic.com/articles/how-to-use-retype-a-tool-that-allows-you-to-use-that-font-that-you-liked-but-impossible-to-know-what-it-is
  • Stephen Coles
    Stephen Coles Posts: 1,007
    for those who regularly use these tools, these limitations are obvious

    I’ve had personal experience that suggests otherwise. Someone passed copywriting to me that was obviously generated by a chat bot. I see plenty of other folks doing the same: using AI as a replacement for creativity. 

    An optimistic view is that it’s usually pretty easy to catch these uses, just as it’s easy to see when someone relies completely on a Photoshop effect as the core concept for their design. Without human creativity and an original concept driving the work, image generators (and now font generators) inevitably seem to produce material that is banal and samey.

  • Ray Larabie
    Ray Larabie Posts: 1,432
    edited October 7
    While I use AI tools in my ad copy process, I don't rely on them to write the copy from scratch, at least not with the current state of AI technology. For most of mine used ChatGPT 3.0 a few years ago. My approach begins with writing detailed ad copy myself, including more information than necessary. This ensures the AI has sufficient accurate data to work with.

    Using carefully crafted prompts from days of experiments, I then have the AI generate ad copy based on my initial drafts. This usually results in factually correct content that aims for the tone I'm seeking. The AI often employs the classic AIDA formula. However, the AI-generated copy isn't perfect. Common issues include confusion between “font” and “typeface,” overuse of “bold” as an adjective, and occasionally adopting a “hard sell” tone. To address these issues, I carefully edit and rewrite the AI's output.

    The result is ad copy that potentially has a better conversion rate than what I might have written entirely on my own. This method allows me to leverage AI's capabilities while maintaining control over the final product, ensuring it aligns with my brand voice and accurately represents my fonts.

    It's important to note that I view good font descriptions and effective font ad copy as distinct entities. While related, they serve different purposes and may require different approaches. By combining my expertise with AI assistance, I aim to create compelling ad copy that effectively showcases my fonts while maintaining accuracy and brand consistency.

    Why not just write it myself? Vocational art high-school, baby! It's a miracle I can even write a sentence.

    @Simon Cozens
    Contrary to popular belief, AI tools aren't ideal for general question-answering. They excel at tasks like coding assistance, writing help, and content summarization. It's crucial to understand their strengths and use them appropriately.

    The “amnesia” associated with AI is actually due to token limits. Claude retains information throughout a conversation, but eventually ends due to resource constraints. ChatGPT, on the other hand, becomes error-prone in longer threads, signaling it's time to start anew. When nearing conversation limits, try reminding the AI of the original instructions and ask it to incorporate key learnings which you can copy/paste for your next conversation. This maximizes efficiency, similar to briefing a new employee. ChatGPT Pro offers additional features for managing longer conversations. But you don't need it. Just keep a document with your ever expanding and improving prompt.

    Remember, these tools are meant to enhance your work, not replace human expertise and judgment. Use them wisely as powerful aids, but don't rely on them for factual information or decision-making. If you expect these tools to do all the thinking, of course, they'll come up with bland content. If you're willing to put in the time experimenting and learning, they can enhance your creativity and productivity.
  • Simon Cozens
    Simon Cozens Posts: 747
    edited October 8
    Ray Larabie said:
    The “amnesia” associated with AI is actually due to token limits.

    "Gell-Mann Amnesia" is a concept originated by Michael Crichton, originally about trusting the press: a scientist (stereotypically, Murray Gell-Mann) can read a newspaper article about physics and realise that 90% of the detail is wrong, but when they turn to the foreign policy pages, they magically forget that the journalism is full of crap and start taking the newspaper seriously again. "Everything seems true apart from that which I have direct experience of."

    I agree that machine learning does best with very specialised models doing very specialised tasks. I use Copilot a lot for coding; although 50% of what it comes up with is complete garbage, it's a good enough autocompletion for monotonous tasks. I'd trust an ML cancer detection or protein folding model.

    But ChatGPT and other LLMs which claim to be more general-purpose are a liability. Yes, of course, use them as a tool. But the problem is not how I use them. I use them just fine. It's how everyone else uses them that's the problem. :smile:

Categories