Google’s Deepmind AI can create ‘3D’ videos based on a single image

A 30-second video can be made from the AI analysing and contextualising the contents of an image

(Image by Alan Warburton / © BBC / Better Images of AI / Nature / CC-BY 4.0)

Google’s Deepmind neural network is now able to create 30-second videos from a single image.

Transframer, as the new tool is called, only needs one photo to work with as it starts identifying what is in the picture’s frame. It analyses the content of the image, and then predicts what would be around it with “context images” – guessing what objects would look like from different angles, based on a huge amount of training data.

“Given a collection of context images with associated annotations (time-stamps, camera viewpoints, etc.), and a query annotation, the task is to predict a probability distribution over the target image,” the Deepmind team wrote in a post.

“This framework supports a range of visual prediction tasks, including video modelling, novel view synthesis, and multi-task vision.”

It is possible that this technology could be used tho build three-dimensional digital environments, rather than the traditional rending that is currently used in video games and other online spaces.

Google’s artificial intelligence team has made other striking developments in the past. Last month, the system mapped out the shape of virtually every protein known to science - a breakthrough that will help to tackle major global challenges such as developing malaria vaccines and fighting plastic pollution.

The division also trained an AI to control the superheated plasma inside a nuclear fusion reactor, opening up new avenues to advance the arrival of unlimited clean energy.

The major challenge of this is shaping and maintaining high-temperature plasma. In stars this is achieved through gravity, but is more challenging on Earth - needing lasers or magnets to control the extremely hot material.

Nevertheless, DeepMind’s AI was able to constantly control the plasma by taking 90 different measurements 10,000 times a second, and adjusting the magnetic field accordingly.