Google’s new AI can restore low-res photos

All the crime dramas out there can finally “enhance” images with software that actually exists. A new artificial intelligence-driven image upscaler from Google can increase a low-resolution 64 x 64 image to a 1024 x 1024 image, with realistic details.  

In a blog post titled “High Fidelity Image Generation Using Diffusion Models”, Google explains how it has developed a pair of AI technologies that can upscale a low-resolution photo by steadily increasing through selective destruction and reconstruction of the original input image. 

Of course, this technology is a reconstruction of what the AI thinks will be there, and may deviate from the original reference image. In its tests, however, it is surprisingly accurate in coming close to the reference image. Watch the video below to see how it works: 

The first component of this AI pair is Super-Resolution via Repeated Refinements (SR3), which applies pure Gaussian noise to a low-resolution image before using noise reduction technology to reconstruct an image that’s unblurred and higher-resolution. 

The second part is using Cascaded Diffusion Models (CDM) to intelligently apply Gaussian noise and blur the output image before going back to step one to start the process again. Google calls this technique “condition augmentation” and is superior to other AI upscaling techniques like BigGAN-deep and VQ-VAE-2. 

For more technical information on the technology, visit Google’s blog here.