Google’s new AI will let you zoom into photos without blurring everything – Mashable SE Asia

It's something we've always wished we had, and now it's almost here.

Google has revealed their work on a new piece of artificial intelligence (AI) tech that will allow users to zoom in on photos or images without them getting blurry. Instead of getting pixelated the photos will continue to look clean and sharp, kind of like the tech used in crime TV series like Crime Scene Investigations.

Something like this, but not quite:

The new AI engines work based on what's known as diffusion models, and involves a technique known as natural image synthesis.

In this technique, details are added to an image that weren't originally captured by the camera or inserted by whatever software was used to make it. The additional details are added using a mix of clever guesswork and deriving info from other similar-looking images.

Of course, this means that the enhanced photo won't turn out exactly like the original, but you can still expect pixelated or blurry images to be transformed into pictures that you can enjoy and use without worrying how others see them.

With regards to the tools used to make this magic happen, Google has said that they currently employ two techniques.

The first is a diffusion model technique called Super-Resolution via Repeated Refinement (or SR3), and basically works by adding extra noise or unpredictable elements to an image, then reversing the process and fixing the image, just like how a photo editor works to improve any image taken from a camera.

Then by using probability math and looking through a huge library of pre-existing images, SR3 can then better predict what a full, clear image should look like despite how blurry or pixelated the original is.

"Diffusion models work by corrupting the training data by progressively adding Gaussian noise, slowly wiping out details in the data until it becomes pure noise, and then training a neural network to reverse this corruption process," explained members of the Google Research team.

The other tool is known as Cascaded Diffusion Models (or CDM), that works by acting as "pipelines" through which diffusion models like SR3 can be directed for the purpose of upgrading images into high-resolution.

While these all sound extremely technical, Google's progress with these techniques have so far been impressive.

In a test conducted on the gigantic image database ImageNet, Google's new AI engine manage to convince 50 volunteer testers that images of human faces generated by SR3 and CDM were real 50 percent of the time pretty amazing for a bunch of images created by performing complex mathematics.

Currently, it's still not known how or when Google plans to introduce their new engine to the mainstream, although the team have confirmed that they're still working on other exciting technologies related to AI and diffusion models.

"With SR3 and CDM, we have pushed the performance of diffusion models to state-of-the-art on super-resolution and class-conditional ImageNet generation benchmarks," the team said. "We are excited to further test the limits of diffusion models for a wide variety of generative modeling problems."

Read more tech stories:

Tamagotchi is back, and this time it's literally clinging to your arm

The pandemic has fundamentally changed the way we find a job

Why you need a secret phone number (and how to get one)

Cover image sourced from Kishor on Unsplash. For illustration purposes only.

Visit link:

Google's new AI will let you zoom into photos without blurring everything - Mashable SE Asia