Pages

Google creates two new technologies CDM and SR3 after diving into the world of diffusion models

Google is known to come up with the most amazing features that are known to change the outlook of the world of technology and yet again the tech giant is breaking through doors with some new AI that they are testing.

The company is playing with a concept called 'diffusion models' and while the company had been working towards it, it wasn't their first priority all these years, but looks like this has changed now. The company on their Google AI Blog posted an article in which they discussed the concept of super resolution images. This basically means giving low resolution and blurred images a more finished and sharp look through the help of AI technology.

The first step and technology in the company's new venture is SR3 technology or in other words Super-Resolution via Repeated Refinement. The technology is based on noise resolution. It takes low resolution images as input and adds noise progressively until only pure noise remains and the image turns into a high resolution image. The feature has the ability to reverse the concept as well by breaking through the noise and returning the image in its usual low resolution form.

The SR3 has proven to do exceptionally well than other form of technologies of the same aspect and showed a confusion rate of 50 percent when used on 8X upscale images, while the other similar existing technologies have only managed to show a rate of 34 percent. This shows how well advanced this technology by Google is and how hard the company has worked towards in creating this so that it can give a more photo realistic effect on images.

But this was not all, after its great outcomes and results of SR3, the company started working towards another feature in the same line of working and introduced its new CDM feature, which is also called class-conditional diffusion model.

The CDM works on the principals of ImageNet, but while ImageNet is a much more complex dataset, CDM is made to be a more convenient use. This technology will help users enhance the size of low resolution imaged in bigger cascade and giving them a clearer look without the pixels breaking. Imaged can be enhanced and upgraded from the sizes of 32 x 32 to 64 x 64, then 256 x 256 and finally the largest size being 1024 x 1024.

Both of the newly introduced technologies are amazing and Google is proud of its team for coming through with such great work and cannot wait to see what else they can accomplish in the world of diffusion models.


Read next: The face filters on social media apps look more realistic when they are applied to the faces they are designed for; however, they do not certainly work on persons of color

No comments:

Post a Comment