Artificial intelligence creates better and better images based on text


Producing artificial intelligence OpenAI company released a year ago software that produced images based on text.

For example, the program created an image of the words “avocado-shaped armchair”.

Now the same company is releasing artificial intelligence that produces even more accurate images even faster and cheaper. The program follows the verbal descriptions of the image more and more closely.

Content will continue after the ad

Artificial intelligence is named Glide. It creates an image in 15 seconds. It has “only” 3.5 billion variables programmed into it.

Content continues below the ad

The previous model of artificial intelligence that produced images from 2021 was called Give her. It contained as many as 12 billion different variables.

The old model learned to make pictures by going through different sets of pictures of dogs, houses and cars, for example.

The new model approaches the production of images in a slightly different way.

Its learning neural network is still being trained by providing a lot of images. At the same time, however, the images displayed are destroyed when the AI ​​is tracked.

A noise layer is applied to the image, which degrades the image quality. In practice, the image is gradually cluttered and destroyed pixel by pixel. The end result is pure chaos.

Artificial intelligence is a learning neural network that learns how chaos is produced. At the same time, however, it can learn to do the opposite.

It can produce a sensible picture based on what you have learned. The beginning of the image is in a few pixels.

Artificial intelligence starts processing an image with a feed that has only noise. This creates a photorealistic image that matches the text description as closely as possible.

Researchers at OpenAI say Glide allows the user to create an image of, for example, a “comfortable living room”.

The user may then ask for more: “Take a closer look at that painting of the dogs above the couch”.

Researcher in Interactive Information Technology, Professor Mark Riedl The Georgia Institute of Technology believes such applications will become part of photo editing.

The image processing program Photoshop can become neural, Riedl estimates In New Scientist.

In the experiments, people were made to monitor and compare the images produced by Glide and Dall-e. Glide’s pictures seemed more realistic.

Open AI researchers presented Glide used by researchers in the arXiv Pre-Release Service. The article has not been peer reviewed.


Source: Tiede by www.tiede.fi.

*The article has been translated based on the content of Tiede by www.tiede.fi. If there is any problem regarding the content, copyright, please leave a report below the article. We will try to process as quickly as possible to protect the rights of the author. Thank you very much!

*We just want readers to access information more quickly and easily with other multilingual content, instead of information only available in a certain language.

*We always respect the copyright of the content of the author and always include the original link of the source article.If the author disagrees, just leave the report below the article, the article will be edited or deleted at the request of the author. Thanks very much! Best regards!