This AI Can Change Weather, Seasons, and Time of Day in Photos

Want to change winter to summer or day to night in a photo? In the future, that could be as easy as a single click. Scientists at NVIDIA are working on a mind-boggling AI system that can change things like seasons, time of day, and weather in photos.

In a new paper titled “Unsupervised Image-to-Image Translation Networks,” the scientists describe how they created a new artificial intelligence framework that allows for unsupervised image-to-image translations.

The scientists write that for many image-to-image translation problems, you can train the AI using high-quality before-and-after examples. For example, to train an AI in turning low-resolution photos into high-resolution ones, you can feed high-res photos and their low-res counterparts into the system to train it.

But for things like changing a rainy scene into a bright sunny one, it’s more difficult to find or create exact before-and-after pairs. Therefore, the AI needs to be trained in an “unsupervised” manner without these precise examples to learn from. The NVIDIA researchers created a new framework for this unsupervised training.

To test the effectiveness of their method, they created a series of results for various image translation tasks.

Given a snowy winter photo, the AI can automatically turn it into a summer one:

…and vice versa:

Given a summer scene, the AI can turn it into a rainy one:

…and vice versa:

Given a daytime scene, the AI can turn it into a nighttime one:

…and vice versa:

Do the conversion on all the frames of a video, and you can change seasons and time of day in those as well:

“We presented a general framework for unsupervised image-to-image translation,” the scientists write. “We showed it learned to translate an image from one domain to another without any corresponding images in two domains in the training dataset.”

If you’re interested in the technical details of how this unsupervised AI training was done, you can read the paper here. You can also find more information on this project on the NVIDIA website.

Discussion