everything about Google’s AI forbidden to the general public

Parti is an artificial intelligence created by Google, capable of generating images from texts. Considered too dangerous for the public, this AI is kept secret. Find out everything you need to know.

Artificial intelligence now makes it possible to generate images from text. After Open AI’s DALL-E, Google launched Imagen based on a similar architecture with a broader AI model.

This tool allows to generate better images from text descriptions through a plus high level of language comprehension.

Now Google just unveiled the new IA Parti model (Autoregressive paths from text to image). This model tested an alternative architecture known as “autoregressive” even closer to the functioning of large language models.

These models predict new words based on precedents, and in the context of the sentence or paragraph. For its part, parti applies this principle to images.

An AI with knowledge of the world

According to Google, party can extend almost unlimited. This is the source of its performance, as all language models outperform better results with full training bringing in more parameters.

This AI can also convert particularly long and complex texts in images, in pictures. She demonstrates a deep understanding of the connection between language and patterns.

Also, Parti can generate images of subjects that weren’t even in its training data. or simply does not exist. Researchers believe that she is capable of providing accurate knowledge of the world, composing many highly detailed characters and objects, and interactions.

She can even respect formats or styles precise images. This AI generated 256×256 pixel definition images. It then uses an upscaler to achieve a 1024×1024 resolution.

party styles

The larger model trained by Google has 20 billion parameters and produces images very close to the texts. According to the firm, he excels at producing drawings from abstract sentences, requiring a rich vocabulary, specific perspectives, writing or symbols.

The human testers preferred products by the largest model in 63% of cases. Furthermore, they showed that this 20 billion parameter model generated images that matched the text in about 76% of cases.

part parameters examples

The models were trained using the Google Cloud TPU, which is able to support the huge number of parameters.

How does party work?

Parti or Pathways Autoregressive Text-to-Image artificial intelligence investigates sets of images called “image tokens” and use them to construct new images.

Tokens and other training materials are the settings, and the realism of the images produced by Parti increased according to the number of parameters. The largest model trained by Google, at 20 billion parameters, generated photorealistic images.

The operation of Part different from that of Imagenthe text-to-image generator designed by Google for diffusion learning. This process involves training the computer by adding “noise” to an image.

The model then learns to noise decoder to recreate the original image. He gradually improves, until he can turn what looks like a series of random dots into an image.

An artificial intelligence too dangerous for the public

Despite his prowess, Party still has limits. She encounters problems to represent the proportions or for the differentiation.

Like DALL-E 2, this AI is unable to count objects on a picture. It may also suffer from technical errors such as one of the colors.

parts failed pictures

The research team also fears that Parti could reproduce biases and stereotypes, to the stage of DALL-E 2 and many other AIs. Stereotypes about typical male and female occupations can be amplified.

Furthermore, this AI could be abused to generate photorealistic Deep Fakes of people and impersonate them. For all these reasons, the researchers made the choice not to publish the modelcode or data for now…

A name in reference to the architecture

The party name is actually a reference to Pathways : the first generation of AI architecture from Google. It was unveiled at the end of 2021 by Jeff Dean, director of AI at Google.

The goal of this versatile AI system is to one day be able to perform millions of different tasks. Everything leads us to believe that party will be used to generate an image within this future architecture.

Several sample images generated by Parti are available on the official website at this address. You will also find detailed explanations of the structure of the system.

Welcome to the era of image generators

Parti and Imagen are not the only models of text-to-images artificial intelligence. In addition to these models created by Google, we can cite OpenAI’s Dall-E, but also VQ-GAN+CLIP and Latent Diffusion Models.

Similarly, the Dall-E Mini tool is an open-source text-to-image AI and accessible to the public. However, it was trained on a smaller data set and does not provide the same level of performance.

Text-to-image AIs are based on GANs or antagonistic neural networks. This type of neural network is based on two algorithms, one of which tries to imitate the training data until it succeeds in fooling the second.

Thanks to GANs, artificial intelligence can also imitate the style of a painter or a musician. In general, this type of neural network allows AI to imitate human artistic creation.

As technology is evaluated, artificial intelligence creates increasingly successful creations. Will she ever be able to surpass the human being?

Leave a Comment