Dol-e artificial intelligence can generate multiple visual interpretations from a single sentence.
OpenAI, an American company specializing in artificial intelligence and machine learning, has launched a new neural network; Slap. The latter can create images from a single title. It may not seem like it, but in reality it is more complicated for a robot.
Give it to her – Abbreviation of the name of the artist Salvador Dali and Wall-E character – Based on the language model of GDP-3, which was created by OpenAI. The latter makes it possible to create original and coherent paragraphs of text from a few words. It has more than 12 billion parameters.
« The GPT-3 model shows that language can be used to instruct a neural network to perform various text generation tasks. The GPT image showed that the same type of neural network could be used to create more reliable images. We extend these results to show that the language manipulation of visual concepts is not yet reached 2, explains the American company.
OpenAI’s new artificial intelligence, on the other hand, is based on a few words or a sentence to create images. AI is an exercise in drawing hundreds of millions of images and their captions into its database to create an image from a short text.
Doll-e not only creates an image, but also provides a lot from the given description. So there is something for all tastes. For example, here are the AI-generated images based on the sentence “ Description of a baby white radish walking a dog »:
The fact that it provides multiple visual interpretations of the same sentence shows that AI is better developed in this precise sense, and that the images do not come from or in the form of Dol-e’s pure creativity. Own intelligence.
« We set it up [Dall-E] Has a variety of skills, including creating human forms of animals and objects, incorporating unrelated ideas, organizing text, and applying transitions to existing images », Soligne OpenIA.
If the US company did not specify the context in which Dol-E would be useful, one can imagine that the latter would be of some sort. It may also serve as a basis for other types of neural networks.