Imagen, Google's AI, creates images from text

You type detailed text, wait a while and then an image never seen on the internet is generated based on the written description. This is how we can explain the Image, new AI of Google that was announced by the company as an interesting option and that should help the industry to understand more about machine learning algorithm (machine learning). 

The novelty joins other artificial intelligence options that already do the same job, but the browser company's advances are greater and the result is really impressive. Check out the first details shared right now. 

Machine learning advanced

Based on artificial intelligence systems, the Google Research, the research arm of the Google, presented the first results that were created – still in the prototype phase – with the Image. When typing text, just wait for the tool to create an image from scratch, without the need to send a graphic reference. The idea is to allow the tool, through artificial intelligence resources, to obtain high-quality photos just using text typing as a basis.

In tests carried out with “competitors” such as the GIVE HER, VQ-WIN+CLIP e Latent Diffusion Model, Google stressed that the Image managed to deliver better results, which puts the tool in the spotlight, at least for now. The work of Google's AI happens in 4 steps, ranging from creating the image to adapting to a higher resolution. 

The results are impressive, as the tests were done with random words like “Teddy bears swimming in the 400m Butterfly Olympics”. Some tests went further and it was possible to see that the developers had fun testing with Imagen, check out some more generated images: 

O Google is one of the companies that has invested the most in artificial intelligence solutions in recent years and just look at the advances in Google Assistant to have a little preview. More than a simple tool that activates or deactivates functions on your smartphone, the personal assistant is already able to make calls in a natural tone, as in some devices, it is no longer necessary to say “Okay Google” to continue a conversation with the assistant. 

During the Google I/O 2022, a novelty that took the name of Pathways Language Model (paLM) was able to solve math problems with text, delivering the result and explanation of how the information was obtained. 

The company expects the paLM can be used in various fields and through machine learning, even learning new languages ​​that are not highly fluent. 

who can use the Image?

For now, the use of Google AI is not released to the public, due to the company having some fear about how the results will be used. As the data to create images was taken from “an uncured web scale”, according to the company itself, there is a risk that text-to-image AI will end up delivering results that may be offensive to some minorities, even more so if asked to creating photos with people. In the published results, there are no images with people, so we don't know yet how this would happen in practice. 

Ao mesmo tempo, o Google Research points out that the training filters were filtered so that inappropriate noise and content, such as pornography, racist insults and social stereotypes, were filtered out. In other words: since the artificial intelligence training was carried out based on what is on the internet, Google fears that the Image be used to create images that arouse anxiety triggers or other problems. 

In the publication that released the AI ​​results, Google also sent a message to its competitors to take the same care when training their artificial intelligences in creating photos with people. 

What did you think of the first results released? tell us Comment! 

See also other features 

Check out all the details about new Google products that were showcased during Google I/O 2022

Sources: Google Research l CNET l Google Blog