Thanks to Fooohx, reader of Korben.info and support on my Twitch, I tested live this week the artificial intelligence DALL-E 2 in beta. For those who missed the phenomenon, DALL-E is an OpenAI project that is capable of producing images from a simple sentence.
Last time, I told you about Craiyon which does the same thing, but with DALL-E, we clearly exceed my expectations in terms of quality. I’m going to spoil you: This thing is crazy!
I first asked him to redo some babies for my site, trying to be as precise as possible. It wasn’t perfect in terms of understanding, but in terms of visual rendering, it’s phew.
Something simpler like “hacker baby with sunglasses” was already much better.
Then we wondered if it was able to generate logos. So I asked him for a logo if Mac Donalds sold shoes. Here’s the result.
Then I wondered if he was able to generate personalities. So Jesus, it’s ok…
Abraham Lincoln is less precise, but we also recognize him.
Dora the Explorer GTA version, it’s also ok…
On the other hand, Gérald Darmanin with a yellow vest, he didn’t tell me the presumed innocent minister, but a whole bunch of guys who have Gérald’s physique dressed in a yellow vest.
At one point, I made a mistake, I also spoke to him in French and DALL-E 2 did almost very well.
In English, the result is still more accurate (pink bear).
One of the aspects that is also the strength of DALL-E is that you can specify styles. You have seen it in my captures, I specify for example “Professional Photo” or “realistic photo”. But you can add stuff like “3D rendering”.
Or Digital Art..
But also artistic styles like pixel art, or in the style of artists like Monet or Keith Haring.
I also tried very utilitarian things like an image for Twitch, album covers or even a military crest.
As you can see, as soon as there is a bit of text, the AI is lost and places letters that are close to the subject, but without really writing correctly. I think this is a major axis that they will have to work on.
Finally, to end, I asked him stuff that requires a little imagination like “What if God were a woman”, or “Mona Lisa a man” or what the head of the Internet looks like, an epic medieval battle between tomatoes and radishes or even a “photo that aspires to happiness and fulfillment”. Here’s the result :
As you can see, there are no limits. The results are sometimes far from what I asked for, but that’s super rare. And each image, if you like, can in turn be declined in other versions automatically. Anyway, in terms of quality, we are still on something super clean. You can see the chasm between DALL-E 2 and Craiyon for example.
Personally, if it remains free, I intend to use it sometimes for my site, in order to have illustration images that stick with my subject and which do not belong to anyone except OpenAI ( you will have to read the T&Cs on this).
And you what do you think ? I am blown away.
Thanks again to everyone who participated in this test live on my Twitch channel and who offered me some of these ideas.