isn’t the only team at Google working on artificial intelligence. There is also Google Brain, which is part of Google AI, which also specializes in deep learning. Researchers from the “Brain Team” have just unveiled Imagen, able to generate photorealistic images from texts.
The project is very similar to, presented just over a month ago. Enter a description of the desired image, even a wacky one like ” a brain riding a rocket heading to the Moon », and the AI will produce the corresponding image.
AI research faces ethical issues
To evaluate Imagen’s performance, the researchers pitted Imagen against DALL-E 2, as well as other similar models with their own evaluation tool DrawBench. They created a series of 200 descriptions, in 11 categories, used on each of the different models. The images obtained were then presented to 25 volunteers for each category, who had to evaluate them according to two criteria: the quality of the image and the correspondence with the description. In both cases, the participants overwhelmingly preferred Imagen.
The tool is not available to the general public for ethical reasons. The researchers have uploaded a simple demonstration that allows you to choose between several options, but nothing more. In addition to the risks of abuse and the possible impacts of such a tool, the researchers wanted to highlight the problem of the databases used, which are made up of unmoderated public images. According to them, ” datasets of this nature often reflectoppressive views and associations that are derogatory or harmful to marginalized groups “.
Read also on Futura