Google’s DeepMind neural system ‘Transframer‘ has shown that it can create short videos from one image frame. It’s pretty impressive to see how it works.
DeepMind tweeted that the “Transframer” artificial intelligence model — a riff of a “transformer from hugging face” type of AI tool — whips up text based upon partial prompts. It can “generate 30 seconds videos from one image.”
Transframer‘s website explains that the AI creates perspective videos by prejudging the surroundings of target images with “context pictures” — basically, it correctly guesses what one or more of the chairs below are from different perspectives. This is based on extensive training data, which allows it to “imagine” an object from another angle.
This model is awe-inspiring as it can apply artificial depth perception to perspective to create what an image would look like if someone “moved” around it. This raises the possibility of whole video games being built on machine learning tech rather than traditional rendering.
Another reason to think about: A Twitter user already stated that he would use Transframer with OpenAI’s DALL-E. This is an excellent example of AI-on-AI action; we’ll see more over the coming years.
Related Popular Stories: