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/ without finding an exact match. Are there other good resources for doing this?
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
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.
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.
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.
This is pretty awesome. I've been working on my side project Mixfont (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.
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!