4 ms·

Show HN: Choosing font combinations with deep learning

121 points by Jack000 9y ago

justswim 9y ago
This is pretty awesome. I've been working on my side project Mixfont (https://www.mixfont.com https://www.mixfont.com) which is a similar sort of font generator. Your pairings are a lot nicer than those on Mixfont though ... would love to learn more about how you did that.
- Jack000 9y ago
  it uses VGG16 to extract font features, then compares those features to figure out the best contrast between fonts. feel free to use the vectors in the Github.
ams6110 9y ago
Tangentially, I've been trying to identify a font. I have a small example as a jpeg image. I tried https://www.myfonts.com/WhatTheFont/ https://www.myfonts.com/WhatTheFont/ without finding an exact match. Are there other good resources for doing this?
- jimhefferon 9y ago
  You could try https://www.reddit.com/r/identifythisfont/ https://www.reddit.com/r/identifythisfont/, and while https://www.reddit.com/r/fontspotting/ https://www.reddit.com/r/fontspotting/ does not seem to be directed toward identifying fonts in the sidebar, it is an active group and there are lots of such questions on the list.
- Jack000 9y ago
  I suspect the tough part of whatthefont isn't so much the visual search but having a large database of labelled fonts. There are other services that does this but personally I've always heard good things about whatthefont.
- teraflop 9y ago
  You might want to try http://www.identifont.com/ http://www.identifont.com/.
- theoh 9y ago
  Can you give us an Imgur link? A lot of people recognize more fonts than you might expect.
LeonidBugaev 9y ago
I wonder, is it possible to generate fonts based on this algorithm, for character sets which are not included into the original font. For example, I have English font, and want exactly same but with Hindy or Cyrillic characters. Any tips on how to implement it?
- Jack000 9y ago
  there's another project that does this: https://erikbern.com/2016/01/21/analyzing-50k-fonts-using-deep-neural-networks.html https://erikbern.com/2016/01/21/analyzing-50k-fonts-using-de... the results are a bit blurry but as the author notes the next step is to use a better loss function. A GAN loss or GAN+L1 ala pix2pix should dramatically improve results.
intoverflow2 9y ago
Interesting, but not all design is about harmony. I'd be more interested in examples where ML could create combinations that shouldn't work together but work. Think like some of the more experimental stuff coming out of the Bloomberg Business Week design team
- danso 9y ago
  That raises an interesting debate about the boundary between experimental-but-brilliant and experimental-but-awful, and whether machine learning can meaningfully discern the difference and (ideally) provide the former. Because much of it is just taste and tastemaking; Bloomberg features, when posted here, often spark comment threads about how ugly or distracting the design is. Usually the majority of commenters side with the designers' apparent intent, but that's because Bloomberg is a relatively well-known/respected organization. Like modern art, experimental design's appeal is inextricably linked to the creator's credentials, e.g. "No one who gets paid that much would do such a crazy thing out of ignorance!". So what would machine learning bring to the mix? I would prefer a heuristics-based analysis. That is, filter out the most popular combinations, and filter out combinations that are linked to "failed" designs (how you measure "failed" would be subjective of course). Then manually select, as a designer, from the uncommon but yet-unhated combinations left over.
  - Declanomous 9y ago
    One of my biggest complaints about things 'curated' via ML is that you are only introduced to things that are relatively similar. It's like trying to find new music by looking at the top 10 songs on the Billboard Chart. Basically ML sucks at curating novelty.
rubatuga 9y ago
Although a title with deep learning in it makes me roll my eyes nowadays, I really should be more open to the ideas behind them. The font pairings at the end of the article look very nice!
- throwanem 9y ago
  I'm looking forward to "Choosing lunch with deep learning", trained on a curated selection of Instagram posts.
  - gjm11 9y ago
    Perhaps this is what you want: https://www.ibmchefwatson.com/ https://www.ibmchefwatson.com/
    - throwanem 9y ago
      Getting tired of living on the worldline where every time I think I'm making a joke, I turn out not to be.
accountyaccount 9y ago
This does an alright job most of the time (7/10 options are decent). That's impressive.
usaphp 9y ago
Honestly most of the results are really poor. I would recommend using https://www.typewolf.com https://www.typewolf.com instead.

121 points by Jack000 9y ago

This is pretty awesome. I've been working on my side project Mixfont (https://www.mixfont.com https://www.mixfont.com) which is a similar sort of font generator. Your pairings are a lot nicer than those on Mixfont though ... would love to learn more about how you did that.

Jack000 9y ago
it uses VGG16 to extract font features, then compares those features to figure out the best contrast between fonts. feel free to use the vectors in the Github.

ams6110 9y ago

Tangentially, I've been trying to identify a font. I have a small example as a jpeg image. I tried https://www.myfonts.com/WhatTheFont/ https://www.myfonts.com/WhatTheFont/ without finding an exact match. Are there other good resources for doing this?

jimhefferon 9y ago
You could try https://www.reddit.com/r/identifythisfont/ https://www.reddit.com/r/identifythisfont/, and while https://www.reddit.com/r/fontspotting/ https://www.reddit.com/r/fontspotting/ does not seem to be directed toward identifying fonts in the sidebar, it is an active group and there are lots of such questions on the list.
Jack000 9y ago
I suspect the tough part of whatthefont isn't so much the visual search but having a large database of labelled fonts. There are other services that does this but personally I've always heard good things about whatthefont.
teraflop 9y ago
You might want to try http://www.identifont.com/ http://www.identifont.com/.
theoh 9y ago
Can you give us an Imgur link? A lot of people recognize more fonts than you might expect.

LeonidBugaev 9y ago

I wonder, is it possible to generate fonts based on this algorithm, for character sets which are not included into the original font. For example, I have English font, and want exactly same but with Hindy or Cyrillic characters. Any tips on how to implement it?

Jack000 9y ago
there's another project that does this: https://erikbern.com/2016/01/21/analyzing-50k-fonts-using-deep-neural-networks.html https://erikbern.com/2016/01/21/analyzing-50k-fonts-using-de... the results are a bit blurry but as the author notes the next step is to use a better loss function. A GAN loss or GAN+L1 ala pix2pix should dramatically improve results.

intoverflow2 9y ago

Interesting, but not all design is about harmony. I'd be more interested in examples where ML could create combinations that shouldn't work together but work. Think like some of the more experimental stuff coming out of the Bloomberg Business Week design team

danso 9y ago
That raises an interesting debate about the boundary between experimental-but-brilliant and experimental-but-awful, and whether machine learning can meaningfully discern the difference and (ideally) provide the former. Because much of it is just taste and tastemaking; Bloomberg features, when posted here, often spark comment threads about how ugly or distracting the design is. Usually the majority of commenters side with the designers' apparent intent, but that's because Bloomberg is a relatively well-known/respected organization. Like modern art, experimental design's appeal is inextricably linked to the creator's credentials, e.g. "No one who gets paid that much would do such a crazy thing out of ignorance!". So what would machine learning bring to the mix? I would prefer a heuristics-based analysis. That is, filter out the most popular combinations, and filter out combinations that are linked to "failed" designs (how you measure "failed" would be subjective of course). Then manually select, as a designer, from the uncommon but yet-unhated combinations left over.
- Declanomous 9y ago
  One of my biggest complaints about things 'curated' via ML is that you are only introduced to things that are relatively similar. It's like trying to find new music by looking at the top 10 songs on the Billboard Chart. Basically ML sucks at curating novelty.

rubatuga 9y ago

Although a title with deep learning in it makes me roll my eyes nowadays, I really should be more open to the ideas behind them. The font pairings at the end of the article look very nice!

throwanem 9y ago
I'm looking forward to "Choosing lunch with deep learning", trained on a curated selection of Instagram posts.
- gjm11 9y ago
  Perhaps this is what you want: https://www.ibmchefwatson.com/ https://www.ibmchefwatson.com/
  - throwanem 9y ago
    Getting tired of living on the worldline where every time I think I'm making a joke, I turn out not to be.

accountyaccount 9y ago

This does an alright job most of the time (7/10 options are decent). That's impressive.

usaphp 9y ago

Honestly most of the results are really poor. I would recommend using https://www.typewolf.com https://www.typewolf.com instead.