Google has made a commitment to Artificial Intelligence integrated into photography for a long time, with the aim of improving images. In its latest novelty we find the ability to convert fully pixelated photos into high resolution photos.
New from Google is an AI-based image scaling technology that improves the quality of low-resolution images. In a Google AI blog post, Brain Team researchers presented two diffusion models to generate high fidelity images. These models are: super resolution imaging (SR3) cascade diffusion (CDM).
SR3 for super-resolution image
The first of these models is the Image Super-Resolution via Repeated Refinement or SR3. This method is defined by the research team as “a super-resolution diffusion model that takes a low-resolution image as input, and builds a high-resolution image from pure noise.”
To do this, the machine uses a process that constantly adds “noise” to an image until only noise remains. When we talk about noise within the world of photography, we are talking about unwanted effects, pixels or arbitrary alteration of brightness and color in an image. With all this, the Google tool totally reverses the process: it eliminates noise and reaches a target distribution “by guiding the input low-resolution image.”
Thanks to this process developed by artificial intelligence, Google can take images with resolutions of 32 x 32 px and can convert them into images of up to 1024 x 1024 px. Remember that Photoshop, the image experts, also has an AI capable of multiplying the resolution of a photo by four and improving the detail.
CDM or the cascade of multiple diffusion models
Once the SR3 proved its worth, the team went further. The researchers added CDM technology. According to the researchers, the CDM is “a class conditional diffusion model trained with ImageNet data to generate high resolution natural images“Google built CDM as” a cascade of multiple broadcast models. ”
As Business Insider explains, this tool starts with a standard diffusion model at the lowest resolution followed by a sequence of super-resolution models They can successively scale the image and add higher resolution details.
With the introduction of these functions, Google aims to improve the natural synthesis of images. Still you have not specified how you will implement these enhancements in our lives as users.
It may be that we find better images within the Google Images search engine or that we can even access these tools to improve our photographs of low quality stored in Google Photos. What the firm has said is that, in addition to daily use, it can also be useful for sectors such as medicine that need high-quality images to perform certain tasks.