AI can generate visuals from text, and eventually it may produce movies.
The company claims that Google's new text-to-image AI outperforms the competitors.
The software, Imagen, delivers a result based on inputted words, such as "a picture of a Persian cat playing the guitar on a beach while wearing a cowboy hat." Imagen is capable of creating both photorealistic and creative renderings.
Other text-to-image generators like DALL-E, VQ-GAN+CLIP, and Latent Diffusion Models are followed by Imagen. People who were asked to examine images made by Imagen and other text-to-image generators discovered that Google's model fared better than rivals in terms of accuracy and visual fidelity.
A lovely corgi lives in a house built of sushi, among other examples of text prompts and the AI's resulting visuals that Google released on its Imagen website. However, these might merely be the finest outcomes produced. Google declined to respond to this story's comments.
Machine learning systems' power is demonstrated through text-to-image learning models. In this instance, Imagen eliminates the requirement for expertise in using specialised software, such as Photoshop, to produce abstract images. AI systems are helping the company come closer to its vision of an ambient computing future, as noted at the Google I/O conference earlier this month. The concept of ambient computing holds that humans will eventually be able to utilise computers naturally without needing to be familiar with specific systems or programming languages.
Google, however, is aware of the potential of text-to-image AI and has decided against making Imagen available to the general public. To learn and produce images, Imagen collects data from the internet. Due of the potential for bias and stereotypes on the internet, Imagen often has these elements. A predilection for lighter skin tones and certain Western gender stereotypes are two examples of the biases, according to Google. The business is concerned that immoral people might misuse Imagen.
According to a white paper released by Google, "Generative approaches present various problems surrounding social and cultural exclusion and bias and can be leveraged for harmful reasons, including harassment and disinformation transmission."
Google warns other AI developers against publishing text-to-image models without carefully considering the data an AI is being trained on.