NVIDIA Research is showing off a new project that uses artificial intelligence to change the time of day and weather in an image. The technology is called “unsupervised image-to-image translation,” and it involves a newly-created framework capable of producing high-quality image translations, such as turning a day photo into a night photo, or a summer photo into a winter photo.
It is, to use the technical term: bananas.
Unsupervised, in this case, refers to a type of AI training that doesn’t involve precise examples upon which the final results can be based. This is due to the variability inherent in taking one type of image, such as one showing a summer day, and translating it into a winter scene. Discussing this, the researchers explained in an abstract:
We compare the proposed framework with competing approaches and present high quality image translation results on various challenging unsupervised image translation tasks, including street scene image translation, animal image translation, and face image translation.
The team provides several before-and-after examples of their AI’s work, demonstrating instances of a sunny day with blue sky being transformed into an overcast day, and a snow-covered winter environment being transformed into a sunny green environment.
The video below shows a video scene transformed from winter to summer:
NVIDIA also shared a video of a day scene transformed in night scene, though the change is far more obvious in this example:
Of course, the transitions are FAR from perfect at this stage, but some of the swaps are so extreme that even the imperfect creations still feel way beyond a computer’s capacity to do by itself. Adobe’s “Deep Fill” and “Project Cloak” are starting to look like a very small taste of the coming AI photo and video editing revolution.