With AI image generators evolving at such a rapid pace, people have been putting them through all kinds of fascinating tests. Someone’s used a new feature in one of the most popular text-to-image tools to see if it could be used to redesign logos – and people even think some of the results are better than the originals.
Midjourney has opened up more possibilities in its platform by incorporating another technology that has been evolving in parallel to image generation: image description. One user has tried using it to describe iconic logos and then use the descriptions as prompts for new designs (see our piece on how to use DALL-E 2 for the basics of AI text-to-image generation).
I asked Midjourney v5 to ‘/describe’ some logos, to see how it would create prompts for them, and to see what it would create in response.Starbucks pic.twitter.com/EiKujhTl1wApril 4, 2023
In addition to using ‘/imagine’ prompts to generate images based on a text description, Midjourney users can now do things the other way around, and reverse engineer images. By uploading an image and using the ‘/describe’ prompt, they can ask Midjourney to create a text description. That can then be tweaked and used, to, in theory, generate similar images with any changes that might be desired.
Since AI image generators don’t produce the same image from the same prompt twice, even using the exact description obtained via ‘/describe’ will generate variations of the original image. A user, who goes by the name of fofrAI on Twitter (opens in new tab), has tried it out on logo designs, from Apple to Starbucks, Twitter and Pepsi (or should that be Pespi?), to see if it could ‘redesign’ their logos for them.
Quite a lot of people seem to think the Starbucks logo redesign is better than the original, although it’s fussier and lacks balance and symmetry. Some of the other designs are just strange. Since AI image generators can’t handle text, Pepsi’s even changed its name.
But I have doubts about the use case. The fact that some of the logos produced are so similar to the real designs seems likely to be a result of their inclusion in Midjourney’s training data rather than the description alone. Type ‘Starbucks logo’ as a prompt into an AI image generator, and you’ll probably find that it knows what that is, although the result might look horrendous.
And while AI image generators can be interesting tools for throwing ideas around, here any successful ideas for redesigns emerge by accident. Even adding specific descriptions to try to tweak the design according to any ideas you might already have in mind is still going to be quite haphazard for now, but we might soon get to a point where you can describe an existing logo and tweak the description to make it simpler or flatter (like we’ve seen so many car logo rebrands) or to tweak the colours and other details.
Read more: