The photo enhancement in movies and TV has shocked people a lot of times. But now, the photo enhancement has jaws dropped by entering into the realm of science fiction. It’s unbelievable to see that high upscaling technology used by Google.
A new article was published on the Google AI blog, titled High Fidelity Image Using Diffusion Model. The researchers of Google shared the breakthrough that they have made in image resolution. This new technology has a machine learning model that can turn any low-resolution image into a high-resolution image. That is beneficial to restore the old family pictures, which get ruined over time. It also includes upgrading medical images.
The concept of diffusion models is in the circle since 2015 but somehow did not get any work on it. Recently, a new model dived into the deep learning models, which is called Deep Generative Models. Google came to know that this new technology has beaten the previous knowledge of technologies when judged by humans.
The super-resolution or the SR3 are the diffusion models. It receives low-resolution images and converts them into high-resolution by using pure noise. The image corruption process is an expert in that. It adds the noise progressively in the image resolution and stops the process when the noise gets pure.
After that, it then reverses the process and removes the pure noise to the input of no noise in low-resolution images. This SR3 process has shown tremendous results on upscaling and natural images. It has an upscaling capability of 8x power with 50% confusion rates. However, the current rate is only 34% and gives photo-realistic results. The blog has shown many photos of changing results from low-resolution images to high-resolution images.
The SR3 has pushed the performance of these models to high satisfaction. The researchers said that they are interested in testing the limitations of these diffusion models. These tests will help to examine a wide range of these generative models.
Written by: Maryem
Reported by: Imaaz Nadeem