The system is trained in interpreting and understanding introductory texts. Imagen then creates a small 64×64 image that fits this interpretation.
Finally, Imagen publishes a series of templates that enlarge the thumbnail to 256 x 256 and then to 1024 x 1024 pixels. Depending on the assignment, the result is a photo-realistic image or a digital oil painting.
Google put the best results online today and the “photos” are usually both funny and captivating. For example, “Extremely Angry Bird” produced an almost photo-realistic bluebird with a cartoonish expression that was out of place for a character from the smartphone game Angry Birds.
Other images, such as “photo of a Corgi dog living in a house made of sushi” or “a chrome duck arguing with an angry turtle in the jungle” have more convincing results — at least until we get into these details. †
‘Imagen works better than Dall-E’
According to the Google team, Imagen would outperform Dall-E, the most successful text-to-image intelligence yet for photorealistic images. Among other things, the team asked people to rate whether the photos were real or taken by Imagen or Dall-E, the winning Imagen photos. For example, Google’s acumen is better able to determine which color descriptions such as “a yellow book and a red vase” belong.
However, there are still significant limitations to such smart programs. For example, Dall-E has a problem with a long and complex introduction. The program also struggles with text that doesn’t perform as expected. seeker Benjamin Hilton for example, he couldn’t correctly flip the two main characters in the painting “The Woman with an Ermine”: if someone is holding a small creature, Dall-E is expecting a baby, not a lady.
Imagen also has problems, the researchers write in the attached document (pdf). The intelligence services did not have a good answer to the phrase “A horse on an astronaut”: all images show an astronaut on horseback. Photorealistic images of people are also very difficult for the generator.
not for the public
For now, Imagen is in the hands of Google researchers. They fear that outsiders will otherwise use the information for perverse purposes such as fake revenge porn and disinformation.
The researchers write that they were shocked to find pornographic material, racist profanity and annoying stereotypes in one of the databases of photos they used to train the AI. First, they want to do more research on how stereotyping and hate speech can affect an intelligence like Imagen in this way.
Source: RTL
Smith Charles is a talented entertainment journalist and author at The Nation View. He has a passion for all things entertainment and writes about the latest movies, TV shows, music, and celebrity news. He’s known for his in-depth interviews with actors, musicians, and other industry figures, and his ability to provide unique perspectives on the entertainment industry.
