It is one of the best known generative AIs, for better or worse. The visual counterpart of ChatGPT, capable of creating oneiric or very realistic images from a few words: Midjourney. The Tech&Co editorial team tested the capabilities and limits of the software during the “IA: le grand test” program at BFM Business. Tests that have some surprises in store.
Available in open beta since summer 2022, Midjourney’s fame has been notably boosted by viral images, such as the fake Pope Francis in a puffer jacket, the fake explosion at the Pentagon broadcast by several very popular Twitter accounts. Or this reinterpretation of the painting the pearl girlexhibited in the museum of the original painting and criticized by many as an “insult” to human art.
Its popularity is also due in large part to its ease of use: once a paid account has been created through the Discord messaging platform, all you have to do is describe in a few words the image you want to appear and the program generates a result. in a few seconds. And even a simple “message” (a request) can give an impressive result. So here is “Pope Francis on the beach in a Hawaiian shirt.”
Then it is possible to add more details: insist on styles (“32k udh”, “vector”, ‘watercolor’), or even more specific words like a particular photographic lens… Each AI comes with a complete vocabulary to know – but which can be changed from top to bottom with the slightest update, and which is totally different depending on the algorithm used: the same indicator will not give the same result at all in Midjourney or other AI like Adobe Firefly.
The results are sometimes… terrifying
Of course, not all results are perfect. Some are even downright terrifying at times… But these abominations are getting fewer and fewer with each update, to the point where it’s sometimes hard to tell the difference at a glance between a fictional image and a real one.
The difficulty is sometimes where you don’t expect it: our attempts to correctly replicate a much-photographed monument, like the Statue of Liberty, failed, no doubt because in such a well-known place, the slightest imperfection is quickly detected.
On the other hand, Midjourney is good at creating images in an advertising style, where the staging and lighting already make the products look “too good to be true”.
Enough for these generative AI to be potentially useful in many contexts: to generate inspiration, to create illustrations, images, logos… Some associations are already using them in their communication campaigns to avoid staging their beneficiaries in humiliating situations.
But these capabilities are not without controversy. Many artists denounce the method used to train these AIs: having them analyze billions of images retrieved from the Internet, without any authorization or compensation from the people behind them, and who then find themselves in competition with these systems.
Source: BFM TV
