Difficult to escape the phenomenon on social networks.
Since the beginning of June, Internet users have fallen in love with the Dall-E Mini site, renamed since Craiyon.
This tool automatically generates a series of nine images from written requests.
The result is often absurd, comical or hilarious.
Some don't hesitate to enter the most unlikely searches, such as "the Little Mermaid eating sushi", "Dora the Explorer at a nuclear test site", or "Donald Trump in Minecraft".
Nearly 50,000 images are now generated every day.
Craiyon is a tribute to the Dall-E project of the OpenAI company.
This is a smaller-scale, open-source replica of this deep neural network capable of creating images on its own from simple text.
Dall-E absorbed huge visual databases and learned what a Labrador looked like, the art deco style or the color yellow.
The visuals created by this artificial intelligence are stunning.
Read alsoDall-E, the AI that is revolutionizing visual creation
“
When OpenAI unveiled its Dall-E project in January 2021, I wanted to try to reproduce the same thing
,” explains
Le Figaro
Boris Dayma, consultant in machine learning.
This Frenchman living in the United States took advantage of a competition launched in July 2021 by the Hugging Face platform and Google to test his idea.
Images automatically generated by Craiyon with the query “portrait of Brigitte Macron in the style of Van Gogh” Craiyon
“
This hackathon provided access to powerful computers and technical support from both companies.
Other people joined me and we ended up winning the competition.
But Dall-E Mini was nowhere near as good as it is today, with low resolution images.
I kept improving the model in my spare time, and it eventually got to a level that made it go viral
” in June 2022. This success led OpenAI to ask Boris Dayma to drop the Dall- E Mini so that there is no more confusion between the two projects.
Approximate results
“
In the beginning, Craiyon was used by artists who work with artificial intelligence.
I would never have believed that it would then be used to create memes on the Internet
, ”smiles this graduate of Centrale Marseille, who has been living in Houston (Texas) for ten years.
To cover the cost of the servers needed to run Craiyon, advertising will appear on the site and paid features are under consideration.
Images automatically generated by Craiyon with the query “diplodocus reading the newspaper, art nouveau style” Craiyon
OpenAI limits access to Dall-E to a hand-picked few thousand testers and has safeguards in place to prevent its technology from being misused.
Craiyon is accessible to everyone.
"No one can doubt that the images created come from software, it is obvious that they are artificial
", continues Boris Dayma.
The AI, for example, does not know how to draw realistic human faces.
“
The model makes rough approximations.
It is not noticeable on a tree, but it is immediately visible on a face.
This technical limitation is ultimately an advantage that makes Craiyon safer.
It's still entertainment.
»
Images automatically generated by Craiyon with the query “Elon Musk in Central Park”.
Crayon
Raising public awareness
Boris Dayma defends the opening of Craiyon to all Internet users.
"
We must now make the general public aware of the existence of this new technology
", which will soon make it possible to create visuals which the human eye will not always be able to suspect that they were designed by a machine, "
in order to develop critical
thinking.
The popular success of Craiyon has effectively shed new media light on the work of OpenAI but also of Google, which is working on a similar project called Imagen.
"
We must also open up the discussion on the limitations and biases of these models
" beyond the circle of researchers in artificial intelligence, continues the designer.
These AIs are trained on large visual databases that have their share of biases.
“
The model is based on what it sees most often.
And so for him, unless he is explicitly told otherwise, a doctor is necessarily a man, and a housekeeper necessarily a woman.
»
Boris Dayma, however, does not believe that such AIs can replace graphic designers and visual designers.
“
These professionals will use these models as a source of inspiration, a way to test ideas and avoid the blank page syndrome.
They will then be able to do much better.
These AIs will be above all tools to help with creation.
»