A 30-second video can be created by AI analyzing the content of an image and placing it in context
Google’s Deepmind neural network can now create 30-second videos from a single image.
The new tool, called Transframer, only needs one photo to work with to start identifying what’s in the image frame. Analyzing the content of the image, Transframer then predicts what will happen around it with “context visuals” – that is, it predicts how objects will look from different angles based on massive amounts of training data.
“Given a stack of context images with associated annotations (timestamps, camera view angles, etc.) and a question description, the task is to predict the probability distribution over the target image,” the Deepmind team wrote in a post.
This system supports a range of visual prediction tasks, including video modeling, new view synthesis and multi-task vision.
This technology could be used instead of traditional image processing, which is already used in video games and other online spaces to create three-dimensional digital environments.
Our team also trained a single model on 8 different tasks across image & videos, e.g., depth estimation, instance segmentation, optical flow, object detection. Learn more via the project website: https://t.co/xywfVPP7b6 2/
— Google DeepMind (@GoogleDeepMind) August 15, 2022
Google’s AI team has made other stunning advances in the past. Last month, it revealed the structure of almost every protein known to science, a breakthrough that will help tackle major global challenges such as developing malaria vaccines and tackling plastic pollution.
Deepmind has opened up new ways to improve access to unlimited clean energy by training an AI to control superheated plasma in a nuclear fusion reactor.
The biggest challenge is shaping the high-temperature plasma and maintaining its state. In stars, this is achieved through gravity, but on Earth it is more difficult; lasers or magnets are needed to control the extremely hot matter.
DeepMind’s AI was able to continuously control the plasma by taking 90 different measurements 10,000 times per second and adjusting the magnetic field accordingly.
the main source of the news: https://www.independent.co.uk/tech
