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


Powered by artificial intelligence that turns short descriptions into visual representations, the platform makes Twitter happy.

A steampunk Pikachu, a Roman Empire Eiffel Tower, or a Roland-Garros-winning Demogorgon: a Twitter account, which lists the AI-generated images Dall-E Mini (or Craiyon), holds so much of a creative paradise than a new circle of the Underworld.

This is because the site, created by two programmers – Pedro Cuenca and Boris Dayma – offers its users the possibility of transforming short descriptions into images. The images are generated in a few seconds by an algorithm, with a realism that is sometimes screaming, sometimes unsuccessful.

The creations posted on Twitter by proud Internet users, for the moment, trace the limits of the platform. He gets along better with cartoon characters, which are easier to represent, or with very simple schemes:

On the other hand, it is more complicated with complex, often deformed faces and shapes. The easiest painters to copy are those with abstract lines, such as the not very comical Francis Bacon, rather than those who adhere to realism.

Concept already used by Google

First published on the recent Unicorn Hugging Face, an American open source platform offering machine learning programs, the program has been republished on its site. And renamed Craiyon, to differentiate itself from the services offered by OpenAI, a pioneer in the sector with its Dall-E service.

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

The example of Tray, the AI ​​developed by Microsoft that became hateful within hours under the influence of machine learning, lingers in the memories. Less spectacularly, the generation of the images is based on stereotyped images, and thus reinforces the clichés (such as, for example, that of a politician on the part of a white man).

The creators of Craiyon also warn of this risk:

While the capabilities of image generation models are impressive, they can also reinforce or exacerbate societal prejudices. Because the model has been trained on unfiltered data from the Internet, it can generate images that contain harmful stereotypes. “

But despite the risks, the models are improving. In an interview with the British media I, Boris Dayma points out that the application proceeds “week after week”. Before other territories are invested according to these algorithms.

There is a video aspect. Basic prototypes already exist – if you can create images, you can stack them to create videos, but they are likely to get 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 willingness to create multimodal searches, and to eventually make it possible to search for any content – image, sound or video – on the Internet. He recently recruited Craiyon developer Boris Dayma into his machine learning program.


Leave a Comment