Dall-E Mini, the site that turns your texts into psychedelic images

Powered by an AI that converts short descriptions into visual representations, the platform makes Twitter happy.

A steampunk Pikachu, an Eiffel Tower from the Roman Empire, or a Demogorgon Roland Garros winner: a Twitter account listing the images generated by the Dall-E Mini (or Craiyon) artificial intelligence thinks so much of a creative paradise ready as a new circle of the underworld.

In fact, the site, created by two programmers – Pedro Cuenca and Boris Dayma – offers its users the possibility of converting short descriptions into images. Visuals are generated in a matter of seconds by an algorithm, with a realism that sometimes screams, sometimes fails.

The creations posted on Twitter by proud netizens are, for the time being, drawing the boundaries of the platform. She does better with cartoon characters, which are easier to represent, or very simple patterns:

On the other hand, it becomes more complicated with faces and complex, often deformed shapes. Easiest to copy are painters with abstract lines, like the very unfunny Francis Bacon, than those who cling to realism.

Concept already used by Google

The program, which was first published on the recently released Unicorn Hugging Face, an American open source platform that offers machine learning programs, has been republished on its own website. And renamed Craiyon to differentiate itself from the services offered by OpenAI, a pioneer in this field with its Dall-E service.

The latter is not available to the general public, just like Imagen, Google’s most recent experience in the field. The Cupertino company justified this unavailability with the need for better moderation of the content on which the artificial intelligence is based: since it learns from the content offered to it on the Internet, the risks of political slippage are great.

The example of Tray, an AI developed by Microsoft, which became hateful in a few hours under the influence of machine learning, is memorable. In a less spectacular way, the image generation is based on stereotypical images, thereby reinforcing clichés (such as that depicting a politician by a white man).

The makers of Craiyon also warn of this risk:

While the skills of image-making models are impressive, they can also reinforce or exacerbate societal prejudice. Because the model was trained with unfiltered data from the web, it can generate images that contain harmful stereotypes.”

But despite the risks, the models are improving. In an interview with British media I, Boris Dayma points out that the application is progressing “week by week”. Before other areas are invested based on these algorithms.

There is a video aspect. There are already basic prototypes – if you can create images, you can stack them to create videos – but this will probably get even better. The next step will be to create 3D objects for games and movies.”

Google is already very active on the subject: the group shows its desire to create multimodal searches and ultimately to enable the search for any content – image, sound or video – on the Internet. He recently recruited Craiyon developer Boris Dayma for his machine learning program.

Leave a Comment