Tech News

Parti, Google’s artificial intelligence to create images from text

This is the new Google AI that generates images based on texts (Europapress)
This is the new Google AI that generates images based on texts (Europapress)

The race to create artificial intelligences that are capable of creating images from text continues, as this is nothing new because in recent weeks several companies have launched their proposals for cloud technologies that can generate content to the public. graphics.

This time the turn is for Google, it has developed its Pathways Autoregressive Text-to-Image model, Parti for its acronym in English, an autoregressive system that creates photorealistic graphic designs from text.

Google is the technology company with the most proposals of this type, for example, last May, it presented “Image”, another artificial intelligence that also creates hyper-realistic images based on texts.

According to the company, these two creations are complementary and furthermore, it suggests that “there are two different families of generative models: diffusion and autoregressive, respectively”. That is, two ways to process the data necessary to create the images.

The interesting thing about Parti is that it is capable of receiving long and complex indications, so its developers made an effort to make this Artificial Intelligence understand the different events that have occurred throughout the history of humanity and show a “precise knowledge of the world” in the words of its creators.

Another objective of this innovation is to include the convergence of different objects and participants to achieve compliance with different formats and styles that are indicated.

In order for Parti to have the ability to understand multiple ideas and concepts, during its development a computer system called PartiPromts (P2) was incorporated, made up of around 1,600 indications in English, both simple and complex.

The foregoing, with the objective that those involved in the project have the possibility of measuring the progress of the tool as it is being used by the public.

It should be noted that Parti can generate images based on descriptions of up to 67 words in length.

This latest artificial intelligence from Google was created by means of big data sequences of images and graphic content, as opposed to the typical data processing of text, numbers and symbols. All these procedures are known as “token sequences” and are carried out through systems that are called Tokenizers, for the case of Parti, the ViT-VQGAN was used.

Image, also from Google, on the other hand, has specialized in short text descriptions but continues to maintain high accuracy, for example, the company claimed that it can also generate images with descriptions based on specific locations or even complex text.

That is, if a user enters “a panther running with a seagull on the beaches of Cancun”, the company’s AI will generate an image based on this description and will understand the scientific name of both animals, as well as the places.

However, it has been mentioned that this AI could generate offensive and/or stereotypical images, so the company has continued to refine it, as well as the new Parti.

Finally, according to Google “Image relies on text encoders trained on uncurated web-scale data, and therefore inherits the social biases and limitations of large linguistic models. As such, there is a risk that Imagen has encoded harmful stereotypes and representations”

:

Related Articles

Back to top button