14 ms·

Hey, computer, make me a font

331 points by pavanyara 3y ago

nitrofurano 3y ago
lots of kernings misfits ftw
- tabtab 3y ago
  Stop your bigotry of Kernians.
dleeftink 3y ago
Although I would be sad to see the handcrafting that goes into designing custom fonts go, some iterations down the line a model like this would greatly aid tedious glyph alignment and consistency tasks when designing CJK, hiragana, katakana and kanji fonts. Inspiring stuff.
- california-og 3y ago
  I think that would be ideal. The 'killer' feature would be: Handcraft a set of control characters, like the letters in "handglove" and then let AI generate the rest. Designing a typeface is fun, until you need to add support for multiple languages and need to make 800+ characters. Or, maybe there is a nice (open source) font, that is unfortunately missing some characters you really need: let AI generate them.
- matsemann 3y ago
  It's already so that writing on computers is quite US-centric (at least English-centric). While this might help on some of the shortcomings, I'm also a bit afraid it will make it so that even more focus is put only on the US part, and the rest of the world get a "good enough" implementation made by AI that kinda erases some heritage.
  - dleeftink 3y ago
    Maybe, but the Latin font market is quite saturated, whereas the CJK space has ample opportunity for innovating and is likely even in need of it, cf. [0][1][2] [0]: https://qz.com/522079/the-long-incredibly-tortuous-and-fascinating-process-of-creating-a-chinese-font https://qz.com/522079/the-long-incredibly-tortuous-and-fasci... [1]: https://fonts.google.com/knowledge/type_in_china_japan_and_korea/japanese_typography_basics https://fonts.google.com/knowledge/type_in_china_japan_and_k... [2]: https://stackoverflow.com/a/14573813 https://stackoverflow.com/a/14573813
logdahl 3y ago
Cool! Now generate 'upper-uppercase' and see what happens :^)
- matsemann 3y ago
  I think this is a reference to "Uppestcase and Lowestcase Letters", a submission a while back about someone training a ML model to generate lowercase/uppercase letters, and used it to uppercase letters already in uppercase. Quite fun https://news.ycombinator.com/item?id=26667852 https://news.ycombinator.com/item?id=26667852
scarygliders 3y ago
Hmmm. The model is a ckpt instead of a safetensor. Pondering on whether to keep proceeding trying this out or not... EDIT: a scan with picklescan[0] found nothing.. exciting. [0] https://github.com/mmaitre314/picklescan https://github.com/mmaitre314/picklescan
- eurekin 3y ago
  proxmox/virtualbox/qemu + throwaway vm
  - scarygliders 3y ago
    Quite, I was thinking about doing so. I just scanned it with picklescan, which found nothing malicious. I just updated my original reply.
- 7moritz7 3y ago
  Haven't seen a single maliscious ckpt file so far. Sure, there is a possibility, but huggingface scans pickled weights automatically so the likelihood of someone using that site to spread malware in this form is super low
  - scarygliders 3y ago
    I've never spotted one in the wild either, but, y'know, I like to not be the one who first finds one out... the bad way. ;)
  - artursapek 3y ago
    “pickled weights”? serious question, how on Earth should someone like me, who has completely missed the last 12 months of AI development, catch up with the state of the art?
    - scarygliders 3y ago
      Just know that the .ckpt format has more or less been replaced by .safetensors these days. tl;dr .ckpt files can contain Python pickles containing runnable Python code, which means a Bad Guy could create a .ckpt model containing malicious python code. Basically.
RugnirViking 3y ago
Ooh I have to try this out when I get home, looks like the weights are under 1GB too
boffinAudio 3y ago
I've long had a project in mind involving the various typefaces of the signage around the city of Vienna, which I find very inspiring in many cases. The idea is to just take a picture of every different typeface I can find, attached to the local buildings at street level. There are some truly wonderful typefaces out there, on signage dating back to last century, and I find the aesthetics often quite appealing. With this tool, could I take a collection of the various typefaces I've captured, and get it to complete the font, such that a sign that only has a few of the required characters could be 'completed' in the same style? Because if so, I'm going to start taking more pictures of Vienna's wonderful types ..
- chris_st 3y ago
  Even if you never get around to using the photos, I think it would be a wonderful service to take the photos and put them up somewhere for non-Vienna residents to enjoy.
  - boffinAudio 3y ago
    Oh, definitely .. but first I must amass an archive worthy of it ..
- simbolit 3y ago
  With this tool: no. With a next-gen tool: if you do some pre-processing on the images, quite possibly.
PaulHoule 3y ago
Kinda funny how it works well at this whereas diffusion models go to die when it comes to drawing text but of course it works in a completely different manner.
- TheRealPomax 3y ago
  There's a huge difference between "pictures of letters" and "writing text" though. Ask stable diffusion to write text and it'll generate hilarious weird-looking results. But, ask it to generate individual letters (e.g. "Show me an ornate uppercase letter b") and it'll do that for you with (mostly) no problems.
- ilaksh 3y ago
  SDXL can do text kind of. Also isn't DALLE-3 a diffusion model? But yeah overall diffusion has not generally been able to do it at all before.
  - gwern 3y ago
    > But yeah overall diffusion has not generally been able to do it at all before. Imagen/Parti were doing text just fine long before DALL-E 3 was announced. GANs were also learning some text in the earlier runup (even ProGAN was doing striking 'moon runes' - amusingly, they were complete gibberish because it did mirroring data augmentation).
lachlan_gray 3y ago
I found a few months ago that the gpt-4 code interpreter is capable of converting a black and white png of a glyph to an svg https://twitter.com/lfegray/status/1678787763905126400 https://twitter.com/lfegray/status/1678787763905126400 It would be cool to combine a script like the one gpt-4 gave me with an image generation model to generate fonts. The approach from this blog post is way more interesting though. On a separate note it reminds me of this suckerpinch video :) maybe we can finally get uppestcase and lowestcase fonts https://www.youtube.com/watch?v=HLRdruqQfRk https://www.youtube.com/watch?v=HLRdruqQfRk
- logicallee 3y ago
  >I found a few months ago that the gpt-4 code interpreter is capable of converting a black and white png of a glyph to an svg :) Easy there, let's not make all the naysayers who say it only just predicts plausible words sweat. Your phrasing almost makes it sound like you're sharing a clear example of it analyzing and completing a complex task correctly, while perfectly understanding what it's doing. Perhaps we should say it only just predicted words that are plausible responses to someone asking to do that, while also predicting plausible words someone might say in response to an error message along the way. It might not actually be doing any converting, just predicting words and tokens without really doing anything. My favorite part of its predictive capabilities is how it is able to predict the other half of a conversation that literally goes "didn't work, try again", "didn't work, try again", "still didn't work, try again", "all right you finally fixed it good job" - without even telling it why it didn't work or quoting the error message. Somehow it is still able to predict the other half of the conversation so that it ends up with "finally, good job!" Who knew that to get results that look like it knows what it's doing, it's enough to predict what could make someone say that! We are truly living in the golden age of statistical prediction that does not involve any degree of thinking, analysis, or understanding. Truly our age of applied statistics is going better than anyone could have, er, "predicted". :)
  - wizzwizz4 3y ago
    > Your phrasing almost makes it sound like you're sharing a clear example of it analyzing and completing a complex task correctly, while perfectly understanding what it's doing. OpenAI has hardcoded (or heavily overfit) several special-purpose functions into their ChatGPT systems. In the past few months, they've integrated other special-purpose models, so their tools can do more than just predictive text (e.g. image recognition). GPT can do limited verbal reasoning, whatever else can do image recognition, but that does not mean the combined system can do visual reasoning. There's no mechanism by which it would (unless you specifically create one, but that's not trivial and doesn't generalise). > Who knew that to get results that look like it knows what it's doing, it's enough to predict what could make someone say that! Everyone. Some call it “specification gaming” or “reward hacking”, and we've known about it for a long time. It's a really obvious concept if you have a good mental model of reinforcement learning. https://doi.org/10.1162%2Fartl_a_00319 https://doi.org/10.1162%2Fartl_a_00319 is a fun example. > We are truly living in the golden age of statistical prediction that does not involve any degree of thinking, analysis, or understanding. This is a straw argument. I can't speak for anyone else, but my criticisms are mainly of people seeing some thinking-like, analysis-like or understanding-like behaviour, and assuming that it is human-like thinking, analysis or understanding, while ignoring other hypotheses (some of which make successful advance predictions in a way the “it's doing what humans do!” models don't). I will note: the people being the most loudly exuberant about ChatGPT's vast intelligence seem to view it as a tool. If I were faced with an opaque box, inside which was a being capable of general-purpose problem solving, conversation, and original thought, my first reaction would not be “I can use this for my own ends”. I am glad that I have seen nothing to convince me that ChatGPT is such a being, and I have theoretical arguments that ChatGPT probably won't ever be such a being, but if you genuinely think this technology has the potential to produce such a being, you have an ethical responsibility.
philipwhiuk 3y ago
> To train the model, I assembled a dataset of 71k distinct fonts. I give it a week before Monotype sues your face off.
- ballenf 3y ago
  Copyright around fonts may not support such a suit in the same as way as works of art. Wikipedia says: "In the United States, the shapes of typefaces are not eligible for copyright but may be protected by design patent (although it is rarely applied for, the first US design patent that was ever awarded was for a typeface).[1]" So just scanning the rendered font (as opposed to the code that generates it), may be harder to stop than scanning of artwork. https://en.wikipedia.org/wiki/Intellectual_property_protection_of_typefaces https://en.wikipedia.org/wiki/Intellectual_property_protecti...
  - jddj 3y ago
    > may be harder to stop than scanning of artwork. Which has not been particularly easy to stop either
  - virtue3 3y ago
    The interesting thing about law is that even if the law doesn't absolutely protect you; the person that you believe infringing on your work for free had better be prepared to pony up lawyer fees to defend their work. I think this would be one of the few times where I think that's useful. Typefaces take a lot of time and consideration and work to create so just blanket ripping off that work because we all take them for granted is kind of bullshit. I have conflicting thoughts about this. But at the end of the day if you only trained on open fonts they just aren't generally as good and the output should be not generally as good as opposed to training on nicer fonts that you technically don't have the rights to (but no one thoughts of this being an issue at the time of design patents / etc ). But we're now in the world where we will pay money to compute an AI model to design fonts instead of just paying designers to design fonts. The race to the bottom is accelerating at an alarming rate.
  - bmitc 3y ago
    General question. What is the end game of all this? Corporations that have the money and other means are just wildly burning electricity while sucking up the world's art and knowledge work. How can anyone possibly protect their work in this no holds barred environment?
dexsst 3y ago
I used to make some fonts for rare, non Latin alphabets like the Orkhon script by hand using a Paint-like freewar, it was fun
scarygliders 3y ago
Okay I can't try it out anyway. "Blocksparse is not available: the current GPU does not expose Tensor cores" My "best" GPU is an RTX 2070 Super, Turing architecture. I've seen similar messages when using stable-diffusion... either with -webui or with automatic, can't exactly remember, but they both run fine on that RTX 2070 Super, so I can only guess that they revert to some other method than Blocksparse on seeing that it doesn't support Turing. Or something. I haven't looked into how they deal with it. I've submitted an Issue [0] for it. I don't have enough knowledge to know if there's some way of saying "don't use Blocksparse" for fontogen. [0] https://github.com/SerCeMan/fontogen/issues/2 https://github.com/SerCeMan/fontogen/issues/2
TheRealPomax 3y ago
Neat! Does it have prompt capabilities for things like FVAR, GSUB, and GPOS? E.g. "okay now include a many-to-one ligature that turns the word 'chicken' into an emoji of a chicken in the same style" or "now make a second, sans-serif, robotic style and add an axis called interpol that varies the font from the style we just made to this new style"?
- simbolit 3y ago
  Not OP, but the answer is "no". What exactly made you suspect such abilities?
  - TheRealPomax 3y ago
    Odd phrasing, but: the part where I've worked on OpenType parsing for decades and love seeing people with a passion for digital typefaces make new and creative tools in that space. Typically folks don't stop working on a cool tool after they write a blog post, they're still refining and extending, so you never know how far someone is trying to take a tool without asking them.
    - simbolit 3y ago
      So, if I understand you correctly, it was less a question "does it do x?" and more an indirect form of "hey OP, would be cool if it did X" ?
      - TheRealPomax 3y ago
        This is not the place for starting a discussion about whether context and subtext should be implicit or explicit in written English. That's what https://philosophy.stackexchange.com https://philosophy.stackexchange.com is for.
        simbolit 3y ago
        I don't want to start a discussion, I wanted to know if I misread your original comment, and whether you meant something different from what I thought you meant. From your answer, tho very indirect, I now suspect that I did misunderstand your initial comment, and answering your question was missing the point. That is all I wanted (after at first wanting to be helpful).
lawlessone 3y ago
This is interesting but i think generating the next letter from the letters before may not be the best way to do it. As you mentioned they degrade with each letter. Maybe creating one long image of a whole font would work better. edit: in the above am misunderstanding what is happening here. But i still think there must be another way to structure this so the attention mechanism doesn't have to work so hard.
- jrmg 3y ago
  Since the first three letters are good, and generated only with the context of the preceding letters, shouldn’t just using the first three (instead of the preceding three) as context for every other one be good enough?
Nevermark 3y ago
In honor of all the times he pressed his hands into his eyes (and myself doing the same thing): I present: “Perplexed” by Nilsa. [0] I have a print in my office, in lieu of a mirror. [0] https://www.sargentsfineart.com/img/nisla/all/nisla-perplexed.jpg https://www.sargentsfineart.com/img/nisla/all/nisla-perplexe...
itsyaboi 3y ago
https://www.youtube.com/watch?v=a8K6QUPmv8Q https://www.youtube.com/watch?v=a8K6QUPmv8Q
paulcnichols 3y ago
Inevitable in a good way. Keep going! There's gold here.
yklcs 3y ago
I’ve tried out some work on generating vector fonts too, in the format of Bezier curves and a seq2seq model. The problem was that fonts outputted by ML models were imprecise. Lines were not perfectly parallel, corners were at 89°, and curves were kinked. It’s not too difficult to get fonts that look good enough, but the imperfections are glaring as fonts are normally perfectly precise. These imperfections are evident in OP’s output too, and in my opinion make these types of models unusable for actual typesetting. A 1% error in a raster output would be pixel colors being slightly off, but a 89° corner in a vector image is immediately noticeable, making this a hard problem to solve. I haven’t looked into this problem too much since, but I’m interested to hear about possible solutions and reading material.
- waterheater 3y ago
  Without changing the fundamental learning process, one could conceivably introduce a "post-production" step, where you tighten up the output according to a set of pre-defined rules (e.g., if an angle is 89 degrees, adjust the angle to 90). Of course, changing the learning process would be best. One idea which comes to mind is finding a way to embed relationships into the ML training system itself (e.g., output no angles other than 90 degrees or some predefined set). Such an approach is a type of contraint-based ML, where the ML agent identifies a solution given certain constraints on the output. In my experience, the right approach to accomplish this goal is using factor graphs.
andybak 3y ago
Everyone knows that AIs can't draw sans...
gigglesupstairs 3y ago
"Fucking Hell" - first thing I yelled to myself when I saw that headline Kudos for the project, of course, but it just saddens me a bit more. Nothing is sacred anymore.
- nvy 3y ago
  >Nothing is sacred anymore. If it can be specified, it can be automated.
  - gigglesupstairs 3y ago
    I know, I know. I am not disputing the technicalities.
- colesantiago 3y ago
  > Kudos for the project, of course, but it just saddens me a bit more. Nothing is sacred anymore. Why does this sadden you? I'm quite happy everything is being done by AI, time will be freed for other things that are more important. Manual font making will not go away though and now anyone can make their own fonts for free.
  - rogerclark 3y ago
    You don't know that the time will be freed for other things that are more important. We don't know for sure how this is all going to work out at all. And people who make fonts, create art, and write prose generally do these things because they like doing them, not because they're forced to. These technologies aren't automating drudgery, they're automating things that give people's lives meaning. What's the endgame here exactly?
    - colesantiago 3y ago
      Time will be freed, ChatGPT, DALL-E, Midjourney and Stable Diffusion has collectively saved countless people billions of hours of time and this will do the same. The big font makers no longer have a hold on extremely pricey fonts that are inaccessible, the general endgame is most software is going free and open source thanks to AI, and that is a good thing.
martincmartin 3y ago
Douglas Hofstader, the author of Godel Escher Bach, thought the task of creating fonts could only be solved with general AI. https://www.m-u-l-t-i-p-l-i-c-i-t-y.org/media/pdf/Metafont-Metamathematics-and-Metaphysics.pdf https://www.m-u-l-t-i-p-l-i-c-i-t-y.org/media/pdf/Metafont-M... The Letter Spirit project aims to model artistic creativity by designing stylistically uniform "gridfonts" (typefaces limited to a grid).
- gwern 3y ago
  I read that a while ago and thought that it was interesting: Hofstadter was right that it would require much more general approaches than Knuth's approach of 'think very hard and tweak a hand-engineered knob', because that's how all the past VAE/GAN/RNN work on typography-related stuff has worked. As for the broader question of whether such approaches are general AI, well, that's a bullet Hofstadter is increasingly willing to bite, as upset as it makes him: https://www.lesswrong.com/posts/kAmgdEjq2eYQkB5PP/douglas-hofstadter-changes-his-mind-on-deep-learning-and-ai https://www.lesswrong.com/posts/kAmgdEjq2eYQkB5PP/douglas-ho...
  - svat 3y ago
    Hofstadter's article is very interesting and delightful (as is typical of him). But as a response to Knuth's article it's basically reacting to a straw-man or misunderstanding: by "a metafont" in "The Concept of a Meta-Font"[1] Knuth simply meant a common description of many related fonts in a family (like the Computer Modern family where different font sizes, bold, italics, sans-serif, typewriter style etc are all generated from common code and tweakable knobs) — this is a consciously chosen and designed family. But when he joked about > The idea of a meta-font should now be clear. But what good is it? The ability to manipulate lots of parameters may be interesting and fun, but does anybody really need a 6⅐-point font that is one fourth of the way between Baskerville and Helvetica? Hofstadter ran with it, imagining Knuth to mean a single universal "metafont" from which every single font can be achieved by suitable tweaking of knobs. This is of course nonsense. Knuth wrote a (little-known or referenced) short response in the same journal's Vol. 17 No. 4 (1983): Volume 17.4 (p 412, or in the PDF page 89 of 96 at https://journals.uc.edu/index.php/vl/issue/view/364/183 https://journals.uc.edu/index.php/vl/issue/view/364/183) [from the tone I imagine him very annoyed :-)]: > I never meant to imply that all typefaces could usefully be combined into one single meta-font, not even if consideration is restricted to book faces. For example, […] Meanwhile, I'm pleased to see that my article has stimulated people to have other ideas, even if those ideas have little or no connection with the main point I was trying to make. Misunderstandings of meta-fonts may well prove to be more important than my own simple observations in the long run. Returning to the thread a bit, all these “write code to draw an image” systems—like Metafont/MetaPost, Asymptote, TikZ (and also I guess DOT/Graphviz, Mermaid, nomnoml, …)—are IMO interesting as a way for those who think in language / symbols / concepts to do visual stuff (and vice-versa to some extent), and also (along Knuth's lines) “truly understand” shapes by translating them into precise descriptions. Metafont was never going to become popular expecting font designers to write code (and the fact that hand-writing SVG is a negligible fraction of usage makes sense), but now that LLMs can help translate back-and-forth, it's going to be interesting to see if we ever get to “understanding” shapes. [1]: https://web.archive.org/web/20220629082019/https://s3-us-west-2.amazonaws.com/visiblelanguage/pdf/16.1/the-concept-of-a-meta-font.pdf https://web.archive.org/web/20220629082019/https://s3-us-wes... / https://journals.uc.edu/index.php/vl/article/view/5329/4193 https://journals.uc.edu/index.php/vl/article/view/5329/4193
euroderf 3y ago
Has anyone tried using an LLM to make a font based on their handwriting ? EDIT: There's a couple (IIRC) of online services that offer this.
- OnlyMortal 3y ago
  If it was my handwriting, it wouldn’t be popular. Perhaps a cursive font might be good though I’m pretty sure one exists. An expert system might be able to join up the letters in cursive and make intentional mistakes to give it the character of natural handwriting?
rogual 3y ago
> THE QUICK BROWN FOX JUMPS OVER THE LAZY DOG It's "a" quick brown fox, otherwise the sentence has no "a".
- lordfrito 3y ago
  Lazy?
- jnosCo 3y ago
  There is an "a" in "lazy"?
  - rogual 3y ago
    Well, I'm stupid.
    - jameshart 3y ago
      The usual mistake people make in reciting this is to say the fox jumped over the lazy dog, causing them to omit an ‘s’ from the sentence. Making sure it’s ‘the’ lazy dog rather than ‘a’ lazy dog is actually important if you care about completing the lowercase alphabet, as without it there’s only an uppercase ‘T’.
  - SamBam 3y ago
    Indeed. But it's worth having "a lazy" anyway to avoid the repetetive "the."
- dave78 3y ago
  Huh, I never gave that sentence much thought, and I guess I never realized it conveniently covered the whole alphabet. It makes so much more sense now!
  - sgerenser 3y ago
    I liked the one that Apple used to use in the MacOS 7/8 era: “How razorback-jumping frogs level six piqued gymnasts!” (Might be slightly off but its unique enough that I remembered it at least 20 years after it disappeared from MacOS).
mastersummoner 3y ago
Poof! You're a font.
Rantenki 3y ago
OK, that's cool, but those fonts are all terrible. The serifs are all different sizes and shapes, sometimes on the same letter. The kerning looks like a random walk. The stroke widths are all over the place, and/or the hinting is busted. Now, that said, it's pretty amazing that this works at all, but it'll take some pretty specific training on a model to get something that can compete with a human made font that's curated for good usability _and_ aesthetics. Sadly, we'll also probably see adoption of these kinds of fonts (along with graphic design, illustration, songwriting, screenwriting, etc)... because "meh, good enough" combined with some Dunning-Kruger. TL;DR: Thanks, I hate it.
- jeron 3y ago
  I don't think any self respecting graphic designer would use these fonts in its current state but it's a cool proof of concept and could be improved upon to a more usable state
- BoorishBears 3y ago
  > Sadly, we'll also probably see adoption of these kinds of fonts (along with graphic design, illustration, songwriting, screenwriting, etc)... because "meh, good enough" combined with some Dunning-Kruger. Ironic bringing up Dunning-Kruger as you treat generic RLHF as a "pretty specific training" and make sweeping declarations about how people will use AI as if the current SOTA of several of the tasks you just mentioned didn't come from not settling for "meh, good enough" and instead applying the "pretty specific training" you alluded to (see Midjourney)
  - Rantenki 3y ago
    I never mentioned reinforcement learning, and my DK statement was completely around using flawed fonts for graphic design, etc. My partner _is_ a professional graphic designer, and we _have_ seen some pretty terrible client graphics that came out of Midjourney. They're amazing for what they are, but it's very difficult to get something out of it that competes with a professional illustrator, even ignoring the whole copyrighted content in the model issue.
kleiba 3y ago
Obligatory xkcd reference: https://xkcd.com/1015/ https://xkcd.com/1015/
Jack000 3y ago
I think this approach isn't ideal because you're representing pixels as 150x150 unique bins. With only 71k fonts it's likely a lot of these bins are never used, especially at the corners. Since you're quantizing anyways, you might as well use a convnet then trace the output, which would better take advantage of the 2d nature of the pixel data. This kind of reminds me of dalle-1 where the image is represented as 256 image tokens then generated one token at a time. That approach is the most direct way to adapt a causal-LM architecture but it clearly didn't make a lot of sense because images don't have a natural top-down-left-right order. For vector graphics, the closest analogous concept to pixel-wise convolution would be the Minkowski sum. I wonder if a Minkowski sum-based diffusion model would work for svg images.
- briandw 3y ago
  How would the Minkowski sum be used in the diffusion model? Is the idea to look at the Minkowski sum of the prediction and label?
  - Jack000 3y ago
    In pixel space a convnet uses pixel-wise convolutions and a pixel-kernel. If you represent a vector image as a polygon, the direct equivalent to a convolution would be the Minkowski sum of the vector image and a polygon-kernel. You could start off with a random polygon and the reverse diffusion process would slowly turn it into a text glyph.
- SerCe 3y ago
  Thank you for the suggestion. A couple of ML engineers with whom I've spoken after publishing the blog also suggested that I should try representing x and y coordinates as separate tokens.
matt3210 3y ago
Granted... You are now a font!
dctoedt 3y ago
"Computer, computer, make me a font | find me a glyph | catch me an X-height" (I'll show myself out.)
fennecfoxy 3y ago
He he the machine learning naysayers gonna jump on this one for sure. Consider a human being designing a scifi styled font; how do they get started? By opening references, of course! To examples of other scifi styled fonts that they do not have the rights to, nor will they credit. Also consider another human being designing a scifi styled font; but instead one that is not allowed to reference the work of anybody else, as some argue machine learning models ought to do. This human being has no references to open, they have not seen any scifi media, be it movies or posters or fonts or anything else. How can they create something like this without any reference at all to it? If a human being creates a scifi font, and their inspiration is not references to other scifi fonts but instead, I don't know, a general concept of the "vibe" they got from watching Blade Runner, must they credit Blade Runner for the inspiration? Must they pay the owner of the Blade Runner rights for their use of ideas from Blade Runner?
acid__ 3y ago
Designing fonts for languages that use Chinese characters is often challenging due to the sheer number of glyphs. This approach to generating fonts is very interesting… feels like it could unlock the creation of heavily stylized fonts that just wouldn’t be feasible otherwise.

331 points by pavanyara 3y ago

nitrofurano 3y ago

lots of kernings misfits ftw

tabtab 3y ago
Stop your bigotry of Kernians.

dleeftink 3y ago

Although I would be sad to see the handcrafting that goes into designing custom fonts go, some iterations down the line a model like this would greatly aid tedious glyph alignment and consistency tasks when designing CJK, hiragana, katakana and kanji fonts. Inspiring stuff.

california-og 3y ago
I think that would be ideal. The 'killer' feature would be: Handcraft a set of control characters, like the letters in "handglove" and then let AI generate the rest. Designing a typeface is fun, until you need to add support for multiple languages and need to make 800+ characters. Or, maybe there is a nice (open source) font, that is unfortunately missing some characters you really need: let AI generate them.
matsemann 3y ago
It's already so that writing on computers is quite US-centric (at least English-centric). While this might help on some of the shortcomings, I'm also a bit afraid it will make it so that even more focus is put only on the US part, and the rest of the world get a "good enough" implementation made by AI that kinda erases some heritage.
- dleeftink 3y ago
  Maybe, but the Latin font market is quite saturated, whereas the CJK space has ample opportunity for innovating and is likely even in need of it, cf. [0][1][2] [0]: https://qz.com/522079/the-long-incredibly-tortuous-and-fascinating-process-of-creating-a-chinese-font https://qz.com/522079/the-long-incredibly-tortuous-and-fasci... [1]: https://fonts.google.com/knowledge/type_in_china_japan_and_korea/japanese_typography_basics https://fonts.google.com/knowledge/type_in_china_japan_and_k... [2]: https://stackoverflow.com/a/14573813 https://stackoverflow.com/a/14573813

logdahl 3y ago

Cool! Now generate 'upper-uppercase' and see what happens :^)

matsemann 3y ago
I think this is a reference to "Uppestcase and Lowestcase Letters", a submission a while back about someone training a ML model to generate lowercase/uppercase letters, and used it to uppercase letters already in uppercase. Quite fun https://news.ycombinator.com/item?id=26667852 https://news.ycombinator.com/item?id=26667852

scarygliders 3y ago

Hmmm. The model is a ckpt instead of a safetensor. Pondering on whether to keep proceeding trying this out or not... EDIT: a scan with picklescan[0] found nothing.. exciting. [0] https://github.com/mmaitre314/picklescan https://github.com/mmaitre314/picklescan

eurekin 3y ago
proxmox/virtualbox/qemu + throwaway vm
- scarygliders 3y ago
  Quite, I was thinking about doing so. I just scanned it with picklescan, which found nothing malicious. I just updated my original reply.
7moritz7 3y ago
Haven't seen a single maliscious ckpt file so far. Sure, there is a possibility, but huggingface scans pickled weights automatically so the likelihood of someone using that site to spread malware in this form is super low
- scarygliders 3y ago
  I've never spotted one in the wild either, but, y'know, I like to not be the one who first finds one out... the bad way. ;)
- artursapek 3y ago
  “pickled weights”? serious question, how on Earth should someone like me, who has completely missed the last 12 months of AI development, catch up with the state of the art?
  - scarygliders 3y ago
    Just know that the .ckpt format has more or less been replaced by .safetensors these days. tl;dr .ckpt files can contain Python pickles containing runnable Python code, which means a Bad Guy could create a .ckpt model containing malicious python code. Basically.

RugnirViking 3y ago

Ooh I have to try this out when I get home, looks like the weights are under 1GB too

boffinAudio 3y ago

I've long had a project in mind involving the various typefaces of the signage around the city of Vienna, which I find very inspiring in many cases. The idea is to just take a picture of every different typeface I can find, attached to the local buildings at street level. There are some truly wonderful typefaces out there, on signage dating back to last century, and I find the aesthetics often quite appealing. With this tool, could I take a collection of the various typefaces I've captured, and get it to complete the font, such that a sign that only has a few of the required characters could be 'completed' in the same style? Because if so, I'm going to start taking more pictures of Vienna's wonderful types ..

chris_st 3y ago
Even if you never get around to using the photos, I think it would be a wonderful service to take the photos and put them up somewhere for non-Vienna residents to enjoy.
- boffinAudio 3y ago
  Oh, definitely .. but first I must amass an archive worthy of it ..
simbolit 3y ago
With this tool: no. With a next-gen tool: if you do some pre-processing on the images, quite possibly.

PaulHoule 3y ago

Kinda funny how it works well at this whereas diffusion models go to die when it comes to drawing text but of course it works in a completely different manner.

TheRealPomax 3y ago
There's a huge difference between "pictures of letters" and "writing text" though. Ask stable diffusion to write text and it'll generate hilarious weird-looking results. But, ask it to generate individual letters (e.g. "Show me an ornate uppercase letter b") and it'll do that for you with (mostly) no problems.
ilaksh 3y ago
SDXL can do text kind of. Also isn't DALLE-3 a diffusion model? But yeah overall diffusion has not generally been able to do it at all before.
- gwern 3y ago
  > But yeah overall diffusion has not generally been able to do it at all before. Imagen/Parti were doing text just fine long before DALL-E 3 was announced. GANs were also learning some text in the earlier runup (even ProGAN was doing striking 'moon runes' - amusingly, they were complete gibberish because it did mirroring data augmentation).

lachlan_gray 3y ago

I found a few months ago that the gpt-4 code interpreter is capable of converting a black and white png of a glyph to an svg https://twitter.com/lfegray/status/1678787763905126400 https://twitter.com/lfegray/status/1678787763905126400 It would be cool to combine a script like the one gpt-4 gave me with an image generation model to generate fonts. The approach from this blog post is way more interesting though. On a separate note it reminds me of this suckerpinch video :) maybe we can finally get uppestcase and lowestcase fonts https://www.youtube.com/watch?v=HLRdruqQfRk https://www.youtube.com/watch?v=HLRdruqQfRk

logicallee 3y ago
>I found a few months ago that the gpt-4 code interpreter is capable of converting a black and white png of a glyph to an svg :) Easy there, let's not make all the naysayers who say it only just predicts plausible words sweat. Your phrasing almost makes it sound like you're sharing a clear example of it analyzing and completing a complex task correctly, while perfectly understanding what it's doing. Perhaps we should say it only just predicted words that are plausible responses to someone asking to do that, while also predicting plausible words someone might say in response to an error message along the way. It might not actually be doing any converting, just predicting words and tokens without really doing anything. My favorite part of its predictive capabilities is how it is able to predict the other half of a conversation that literally goes "didn't work, try again", "didn't work, try again", "still didn't work, try again", "all right you finally fixed it good job" - without even telling it why it didn't work or quoting the error message. Somehow it is still able to predict the other half of the conversation so that it ends up with "finally, good job!" Who knew that to get results that look like it knows what it's doing, it's enough to predict what could make someone say that! We are truly living in the golden age of statistical prediction that does not involve any degree of thinking, analysis, or understanding. Truly our age of applied statistics is going better than anyone could have, er, "predicted". :)
- wizzwizz4 3y ago
  > Your phrasing almost makes it sound like you're sharing a clear example of it analyzing and completing a complex task correctly, while perfectly understanding what it's doing. OpenAI has hardcoded (or heavily overfit) several special-purpose functions into their ChatGPT systems. In the past few months, they've integrated other special-purpose models, so their tools can do more than just predictive text (e.g. image recognition). GPT can do limited verbal reasoning, whatever else can do image recognition, but that does not mean the combined system can do visual reasoning. There's no mechanism by which it would (unless you specifically create one, but that's not trivial and doesn't generalise). > Who knew that to get results that look like it knows what it's doing, it's enough to predict what could make someone say that! Everyone. Some call it “specification gaming” or “reward hacking”, and we've known about it for a long time. It's a really obvious concept if you have a good mental model of reinforcement learning. https://doi.org/10.1162%2Fartl_a_00319 https://doi.org/10.1162%2Fartl_a_00319 is a fun example. > We are truly living in the golden age of statistical prediction that does not involve any degree of thinking, analysis, or understanding. This is a straw argument. I can't speak for anyone else, but my criticisms are mainly of people seeing some thinking-like, analysis-like or understanding-like behaviour, and assuming that it is human-like thinking, analysis or understanding, while ignoring other hypotheses (some of which make successful advance predictions in a way the “it's doing what humans do!” models don't). I will note: the people being the most loudly exuberant about ChatGPT's vast intelligence seem to view it as a tool. If I were faced with an opaque box, inside which was a being capable of general-purpose problem solving, conversation, and original thought, my first reaction would not be “I can use this for my own ends”. I am glad that I have seen nothing to convince me that ChatGPT is such a being, and I have theoretical arguments that ChatGPT probably won't ever be such a being, but if you genuinely think this technology has the potential to produce such a being, you have an ethical responsibility.

philipwhiuk 3y ago

> To train the model, I assembled a dataset of 71k distinct fonts. I give it a week before Monotype sues your face off.

ballenf 3y ago
Copyright around fonts may not support such a suit in the same as way as works of art. Wikipedia says: "In the United States, the shapes of typefaces are not eligible for copyright but may be protected by design patent (although it is rarely applied for, the first US design patent that was ever awarded was for a typeface).[1]" So just scanning the rendered font (as opposed to the code that generates it), may be harder to stop than scanning of artwork. https://en.wikipedia.org/wiki/Intellectual_property_protection_of_typefaces https://en.wikipedia.org/wiki/Intellectual_property_protecti...
- jddj 3y ago
  > may be harder to stop than scanning of artwork. Which has not been particularly easy to stop either
- virtue3 3y ago
  The interesting thing about law is that even if the law doesn't absolutely protect you; the person that you believe infringing on your work for free had better be prepared to pony up lawyer fees to defend their work. I think this would be one of the few times where I think that's useful. Typefaces take a lot of time and consideration and work to create so just blanket ripping off that work because we all take them for granted is kind of bullshit. I have conflicting thoughts about this. But at the end of the day if you only trained on open fonts they just aren't generally as good and the output should be not generally as good as opposed to training on nicer fonts that you technically don't have the rights to (but no one thoughts of this being an issue at the time of design patents / etc ). But we're now in the world where we will pay money to compute an AI model to design fonts instead of just paying designers to design fonts. The race to the bottom is accelerating at an alarming rate.
- bmitc 3y ago
  General question. What is the end game of all this? Corporations that have the money and other means are just wildly burning electricity while sucking up the world's art and knowledge work. How can anyone possibly protect their work in this no holds barred environment?

dexsst 3y ago

I used to make some fonts for rare, non Latin alphabets like the Orkhon script by hand using a Paint-like freewar, it was fun

scarygliders 3y ago

Okay I can't try it out anyway. "Blocksparse is not available: the current GPU does not expose Tensor cores" My "best" GPU is an RTX 2070 Super, Turing architecture. I've seen similar messages when using stable-diffusion... either with -webui or with automatic, can't exactly remember, but they both run fine on that RTX 2070 Super, so I can only guess that they revert to some other method than Blocksparse on seeing that it doesn't support Turing. Or something. I haven't looked into how they deal with it. I've submitted an Issue [0] for it. I don't have enough knowledge to know if there's some way of saying "don't use Blocksparse" for fontogen. [0] https://github.com/SerCeMan/fontogen/issues/2 https://github.com/SerCeMan/fontogen/issues/2

TheRealPomax 3y ago

Neat! Does it have prompt capabilities for things like FVAR, GSUB, and GPOS? E.g. "okay now include a many-to-one ligature that turns the word 'chicken' into an emoji of a chicken in the same style" or "now make a second, sans-serif, robotic style and add an axis called interpol that varies the font from the style we just made to this new style"?

simbolit 3y ago
Not OP, but the answer is "no". What exactly made you suspect such abilities?
- TheRealPomax 3y ago
  Odd phrasing, but: the part where I've worked on OpenType parsing for decades and love seeing people with a passion for digital typefaces make new and creative tools in that space. Typically folks don't stop working on a cool tool after they write a blog post, they're still refining and extending, so you never know how far someone is trying to take a tool without asking them.
  - simbolit 3y ago
    So, if I understand you correctly, it was less a question "does it do x?" and more an indirect form of "hey OP, would be cool if it did X" ?
    - TheRealPomax 3y ago
      This is not the place for starting a discussion about whether context and subtext should be implicit or explicit in written English. That's what https://philosophy.stackexchange.com https://philosophy.stackexchange.com is for.
      - simbolit 3y ago
        I don't want to start a discussion, I wanted to know if I misread your original comment, and whether you meant something different from what I thought you meant. From your answer, tho very indirect, I now suspect that I did misunderstand your initial comment, and answering your question was missing the point. That is all I wanted (after at first wanting to be helpful).

lawlessone 3y ago

This is interesting but i think generating the next letter from the letters before may not be the best way to do it. As you mentioned they degrade with each letter. Maybe creating one long image of a whole font would work better. edit: in the above am misunderstanding what is happening here. But i still think there must be another way to structure this so the attention mechanism doesn't have to work so hard.

jrmg 3y ago
Since the first three letters are good, and generated only with the context of the preceding letters, shouldn’t just using the first three (instead of the preceding three) as context for every other one be good enough?

Nevermark 3y ago

In honor of all the times he pressed his hands into his eyes (and myself doing the same thing): I present: “Perplexed” by Nilsa. [0] I have a print in my office, in lieu of a mirror. [0] https://www.sargentsfineart.com/img/nisla/all/nisla-perplexed.jpg https://www.sargentsfineart.com/img/nisla/all/nisla-perplexe...

itsyaboi 3y ago

https://www.youtube.com/watch?v=a8K6QUPmv8Q https://www.youtube.com/watch?v=a8K6QUPmv8Q

paulcnichols 3y ago

Inevitable in a good way. Keep going! There's gold here.

yklcs 3y ago

I’ve tried out some work on generating vector fonts too, in the format of Bezier curves and a seq2seq model. The problem was that fonts outputted by ML models were imprecise. Lines were not perfectly parallel, corners were at 89°, and curves were kinked. It’s not too difficult to get fonts that look good enough, but the imperfections are glaring as fonts are normally perfectly precise. These imperfections are evident in OP’s output too, and in my opinion make these types of models unusable for actual typesetting. A 1% error in a raster output would be pixel colors being slightly off, but a 89° corner in a vector image is immediately noticeable, making this a hard problem to solve. I haven’t looked into this problem too much since, but I’m interested to hear about possible solutions and reading material.

waterheater 3y ago
Without changing the fundamental learning process, one could conceivably introduce a "post-production" step, where you tighten up the output according to a set of pre-defined rules (e.g., if an angle is 89 degrees, adjust the angle to 90). Of course, changing the learning process would be best. One idea which comes to mind is finding a way to embed relationships into the ML training system itself (e.g., output no angles other than 90 degrees or some predefined set). Such an approach is a type of contraint-based ML, where the ML agent identifies a solution given certain constraints on the output. In my experience, the right approach to accomplish this goal is using factor graphs.

andybak 3y ago

Everyone knows that AIs can't draw sans...

gigglesupstairs 3y ago

"Fucking Hell" - first thing I yelled to myself when I saw that headline Kudos for the project, of course, but it just saddens me a bit more. Nothing is sacred anymore.

nvy 3y ago
>Nothing is sacred anymore. If it can be specified, it can be automated.
- gigglesupstairs 3y ago
  I know, I know. I am not disputing the technicalities.
colesantiago 3y ago
> Kudos for the project, of course, but it just saddens me a bit more. Nothing is sacred anymore. Why does this sadden you? I'm quite happy everything is being done by AI, time will be freed for other things that are more important. Manual font making will not go away though and now anyone can make their own fonts for free.
- rogerclark 3y ago
  You don't know that the time will be freed for other things that are more important. We don't know for sure how this is all going to work out at all. And people who make fonts, create art, and write prose generally do these things because they like doing them, not because they're forced to. These technologies aren't automating drudgery, they're automating things that give people's lives meaning. What's the endgame here exactly?
  - colesantiago 3y ago
    Time will be freed, ChatGPT, DALL-E, Midjourney and Stable Diffusion has collectively saved countless people billions of hours of time and this will do the same. The big font makers no longer have a hold on extremely pricey fonts that are inaccessible, the general endgame is most software is going free and open source thanks to AI, and that is a good thing.

martincmartin 3y ago

Douglas Hofstader, the author of Godel Escher Bach, thought the task of creating fonts could only be solved with general AI. https://www.m-u-l-t-i-p-l-i-c-i-t-y.org/media/pdf/Metafont-Metamathematics-and-Metaphysics.pdf https://www.m-u-l-t-i-p-l-i-c-i-t-y.org/media/pdf/Metafont-M... The Letter Spirit project aims to model artistic creativity by designing stylistically uniform "gridfonts" (typefaces limited to a grid).

gwern 3y ago
I read that a while ago and thought that it was interesting: Hofstadter was right that it would require much more general approaches than Knuth's approach of 'think very hard and tweak a hand-engineered knob', because that's how all the past VAE/GAN/RNN work on typography-related stuff has worked. As for the broader question of whether such approaches are general AI, well, that's a bullet Hofstadter is increasingly willing to bite, as upset as it makes him: https://www.lesswrong.com/posts/kAmgdEjq2eYQkB5PP/douglas-hofstadter-changes-his-mind-on-deep-learning-and-ai https://www.lesswrong.com/posts/kAmgdEjq2eYQkB5PP/douglas-ho...
- svat 3y ago
  Hofstadter's article is very interesting and delightful (as is typical of him). But as a response to Knuth's article it's basically reacting to a straw-man or misunderstanding: by "a metafont" in "The Concept of a Meta-Font"[1] Knuth simply meant a common description of many related fonts in a family (like the Computer Modern family where different font sizes, bold, italics, sans-serif, typewriter style etc are all generated from common code and tweakable knobs) — this is a consciously chosen and designed family. But when he joked about > The idea of a meta-font should now be clear. But what good is it? The ability to manipulate lots of parameters may be interesting and fun, but does anybody really need a 6⅐-point font that is one fourth of the way between Baskerville and Helvetica? Hofstadter ran with it, imagining Knuth to mean a single universal "metafont" from which every single font can be achieved by suitable tweaking of knobs. This is of course nonsense. Knuth wrote a (little-known or referenced) short response in the same journal's Vol. 17 No. 4 (1983): Volume 17.4 (p 412, or in the PDF page 89 of 96 at https://journals.uc.edu/index.php/vl/issue/view/364/183 https://journals.uc.edu/index.php/vl/issue/view/364/183) [from the tone I imagine him very annoyed :-)]: > I never meant to imply that all typefaces could usefully be combined into one single meta-font, not even if consideration is restricted to book faces. For example, […] Meanwhile, I'm pleased to see that my article has stimulated people to have other ideas, even if those ideas have little or no connection with the main point I was trying to make. Misunderstandings of meta-fonts may well prove to be more important than my own simple observations in the long run. Returning to the thread a bit, all these “write code to draw an image” systems—like Metafont/MetaPost, Asymptote, TikZ (and also I guess DOT/Graphviz, Mermaid, nomnoml, …)—are IMO interesting as a way for those who think in language / symbols / concepts to do visual stuff (and vice-versa to some extent), and also (along Knuth's lines) “truly understand” shapes by translating them into precise descriptions. Metafont was never going to become popular expecting font designers to write code (and the fact that hand-writing SVG is a negligible fraction of usage makes sense), but now that LLMs can help translate back-and-forth, it's going to be interesting to see if we ever get to “understanding” shapes. [1]: https://web.archive.org/web/20220629082019/https://s3-us-west-2.amazonaws.com/visiblelanguage/pdf/16.1/the-concept-of-a-meta-font.pdf https://web.archive.org/web/20220629082019/https://s3-us-wes... / https://journals.uc.edu/index.php/vl/article/view/5329/4193 https://journals.uc.edu/index.php/vl/article/view/5329/4193

euroderf 3y ago

Has anyone tried using an LLM to make a font based on their handwriting ? EDIT: There's a couple (IIRC) of online services that offer this.

OnlyMortal 3y ago
If it was my handwriting, it wouldn’t be popular. Perhaps a cursive font might be good though I’m pretty sure one exists. An expert system might be able to join up the letters in cursive and make intentional mistakes to give it the character of natural handwriting?

rogual 3y ago

> THE QUICK BROWN FOX JUMPS OVER THE LAZY DOG It's "a" quick brown fox, otherwise the sentence has no "a".

lordfrito 3y ago
Lazy?
jnosCo 3y ago
There is an "a" in "lazy"?
- rogual 3y ago
  Well, I'm stupid.
  - jameshart 3y ago
    The usual mistake people make in reciting this is to say the fox jumped over the lazy dog, causing them to omit an ‘s’ from the sentence. Making sure it’s ‘the’ lazy dog rather than ‘a’ lazy dog is actually important if you care about completing the lowercase alphabet, as without it there’s only an uppercase ‘T’.
- SamBam 3y ago
  Indeed. But it's worth having "a lazy" anyway to avoid the repetetive "the."
dave78 3y ago
Huh, I never gave that sentence much thought, and I guess I never realized it conveniently covered the whole alphabet. It makes so much more sense now!
- sgerenser 3y ago
  I liked the one that Apple used to use in the MacOS 7/8 era: “How razorback-jumping frogs level six piqued gymnasts!” (Might be slightly off but its unique enough that I remembered it at least 20 years after it disappeared from MacOS).

mastersummoner 3y ago

Poof! You're a font.

Rantenki 3y ago

OK, that's cool, but those fonts are all terrible. The serifs are all different sizes and shapes, sometimes on the same letter. The kerning looks like a random walk. The stroke widths are all over the place, and/or the hinting is busted. Now, that said, it's pretty amazing that this works at all, but it'll take some pretty specific training on a model to get something that can compete with a human made font that's curated for good usability _and_ aesthetics. Sadly, we'll also probably see adoption of these kinds of fonts (along with graphic design, illustration, songwriting, screenwriting, etc)... because "meh, good enough" combined with some Dunning-Kruger. TL;DR: Thanks, I hate it.

jeron 3y ago
I don't think any self respecting graphic designer would use these fonts in its current state but it's a cool proof of concept and could be improved upon to a more usable state
BoorishBears 3y ago
> Sadly, we'll also probably see adoption of these kinds of fonts (along with graphic design, illustration, songwriting, screenwriting, etc)... because "meh, good enough" combined with some Dunning-Kruger. Ironic bringing up Dunning-Kruger as you treat generic RLHF as a "pretty specific training" and make sweeping declarations about how people will use AI as if the current SOTA of several of the tasks you just mentioned didn't come from not settling for "meh, good enough" and instead applying the "pretty specific training" you alluded to (see Midjourney)
- Rantenki 3y ago
  I never mentioned reinforcement learning, and my DK statement was completely around using flawed fonts for graphic design, etc. My partner _is_ a professional graphic designer, and we _have_ seen some pretty terrible client graphics that came out of Midjourney. They're amazing for what they are, but it's very difficult to get something out of it that competes with a professional illustrator, even ignoring the whole copyrighted content in the model issue.

kleiba 3y ago

Obligatory xkcd reference: https://xkcd.com/1015/ https://xkcd.com/1015/

Jack000 3y ago

I think this approach isn't ideal because you're representing pixels as 150x150 unique bins. With only 71k fonts it's likely a lot of these bins are never used, especially at the corners. Since you're quantizing anyways, you might as well use a convnet then trace the output, which would better take advantage of the 2d nature of the pixel data. This kind of reminds me of dalle-1 where the image is represented as 256 image tokens then generated one token at a time. That approach is the most direct way to adapt a causal-LM architecture but it clearly didn't make a lot of sense because images don't have a natural top-down-left-right order. For vector graphics, the closest analogous concept to pixel-wise convolution would be the Minkowski sum. I wonder if a Minkowski sum-based diffusion model would work for svg images.

briandw 3y ago
How would the Minkowski sum be used in the diffusion model? Is the idea to look at the Minkowski sum of the prediction and label?
- Jack000 3y ago
  In pixel space a convnet uses pixel-wise convolutions and a pixel-kernel. If you represent a vector image as a polygon, the direct equivalent to a convolution would be the Minkowski sum of the vector image and a polygon-kernel. You could start off with a random polygon and the reverse diffusion process would slowly turn it into a text glyph.
SerCe 3y ago
Thank you for the suggestion. A couple of ML engineers with whom I've spoken after publishing the blog also suggested that I should try representing x and y coordinates as separate tokens.

matt3210 3y ago

Granted... You are now a font!

dctoedt 3y ago

"Computer, computer, make me a font | find me a glyph | catch me an X-height" (I'll show myself out.)

fennecfoxy 3y ago

He he the machine learning naysayers gonna jump on this one for sure. Consider a human being designing a scifi styled font; how do they get started? By opening references, of course! To examples of other scifi styled fonts that they do not have the rights to, nor will they credit. Also consider another human being designing a scifi styled font; but instead one that is not allowed to reference the work of anybody else, as some argue machine learning models ought to do. This human being has no references to open, they have not seen any scifi media, be it movies or posters or fonts or anything else. How can they create something like this without any reference at all to it? If a human being creates a scifi font, and their inspiration is not references to other scifi fonts but instead, I don't know, a general concept of the "vibe" they got from watching Blade Runner, must they credit Blade Runner for the inspiration? Must they pay the owner of the Blade Runner rights for their use of ideas from Blade Runner?

acid__ 3y ago

Designing fonts for languages that use Chinese characters is often challenging due to the sheer number of glyphs. This approach to generating fonts is very interesting… feels like it could unlock the creation of heavily stylized fonts that just wouldn’t be feasible otherwise.