We tried designing a logo with AI. The results were… interesting.
We had just finished designing a new identity for a company and were about to present it, when it occurred to us to try again using AI. Would we get a better result?
AI-generated imagery has come a LONG way in the last couple of years. Take ‘Dall.e.2‘ for example, Openai’s foray into this space. Certain inputs, such as ‘A painting of a cat wearing a hat in the style of Vincent Van Gough‘ spits out a rather impressive – and convincing – set of results:
It’s a little less impressive when it comes to photographs, however. Here’s what we got when we inputted: ‘A photograph of a cat wearing a hat in a field of sunflowers‘ (sticking with the Van Gough theme):
Clearly, there’s more work to do to get photo-real images, but oddly enough, some more abstract requests are much better. Here’s ‘A lamp in the shape of a piece of toast‘:
So how would it fare when asked to create a logo?
To keep things simple, we needed to create a logo for a company called ‘Sabre Financial’, based on the idea of ‘cutting through the BS’ of getting a mortgage. Their name, ‘Sabre’, originated from this idea, and we very simply needed to bring this to life.
This is bread and butter stuff for designers, so it wasn’t long before we had a few concepts, eventually whittling it down to one we were happy to present:
So, with all that in mind, we took a stab at generating our first AI logo.
Attempt 1: ‘A logo for a company called Sabre Financial’
This one was a Hail Mary really, just to see what we’d get. The results were, as you might expect, not right:
Straight away, for some unthinkable reason, it altered the company name, preferring instead… ‘sabbare’? or ‘Sabaabre’? Odd, but OK.
We also seem to have various incarnations of ‘S’ as part of the logo, which isn’t a bad starting point, but what the AI wasn’t doing – which a designer would – was taking the noun, ‘sabre’, and conceptualising that as cutting, or swinging, etc.
Clearly, we’d need to go into more specific detail to get anything interesting.
Attempt 2: ‘A house being cut in half by a sword, as a logo’
Having got next to zero conceptual thinking in attempt 1, we realised we’d need to get more specific with our inputs. So we started with something based on our own idea, and in line with the client’s brief.
So we asked the AI to give us ‘a house being cut in half by a sword, as a logo‘, similar to our chosen idea.
Here’s what we got:
Now what’s interesting here is that the AI has again created a name for the company, recognising that a name is somehow involved in a ‘logo’, despite us not really asking for it.
Moreover, we actually thought one of these ideas wasn’t too bad. At least now we were getting semi-reasonable graphics. So we asked for more variations of the second idea, and here’s what we got:
There’s a vague similarity to all of these now, but nothing better than the original. For the briefest of moments, we thought the AI would be able to riff on this variation and come up with something interesting, but alas, it was time to try something different.
Attempt 3: Using our concept as a starting point.
The AI can also take in an uploaded image, and provide variations on that. So we decided we’d offer up our own logo as a starting point, to see if the AI could do better. As a reminder, here’s our graphic, sans the company name:
The AI then gave us these variations:
OK so again, some have weird, extraneous angles and cuts, but you can see the logic behind some of them; in variation 1, it seems to think our house graphic is some kind of map marker, and has added the pointy thing at the bottom. In variation 3, it’s added some kind of roof. The cuts are all pretty reasonable as well, though not quite as clean and nice as ours.
What was becoming evident at this stage of the experiment was that AI had gotten good at certain types of execution, but remained poor as an incubator of ideas.
A ‘human’ designer, for example, recognises intuitively that a cutting motion might carry material with it through the cut, which is why there is a ‘follow through’ of sorts from right-to-left on our original. This helps gives the sense that something has taken a swipe in that direction.
Final attempt: ‘a logo for a finance company’
As a last-ditch attempt, we went back to the most basic input, figuring it might give us a starting point to work with.
But no, it didn’t:
Conclusion: AI can’t make logos… yet.
Although impressive, these Ai generated images have a long way to go to match the creative output of experienced designers. But if we had to bet, we’d say it won’t be long before they can. The limitation at present seems to be that the outputs don’t utilise any creativity or subtle intuition; ask for a sword cutting through a house, and you get a sword and a house, not necessarily doing the cutting.
Some of the early results with cats, however, were remarkably good, and given time I can see the cheaper freelance sites being supplanted for this type of work.
But when all is said and done, nothing (so far) can match a human’s ability to marry abstract concepts and bring them to life visually. So the question remains – will AI prove to be a great designer, and if it does, are we human designers ready for the disruption?