Google’s AI-powered image upscaling technology can generate high-resolution images from low-resolution images.

Google has announced the release of AI-based image upscaling technology, which the company claims will improve the clarity of low-resolution images. The researchers from Google’s Brain Team published a post on the company’s AI blog in which they described two diffusion models for generating high-fidelity images. Models of image super-resolution (SR3) and cascaded diffusion are the two types of models (CDM).

What is Image Super-Resolution and how does it work?

The picture Super-Resolution by Repeated Refinement, often known as SR3, is the first component of this concept. “A super-resolution diffusion model that takes as input a low-resolution image and generates as output a corresponding high-resolution image from pure noise,” according to the research team.

Google’s research team has demonstrated some amazing results, which demonstrate how this technology may be utilised to effectively improve the image quality of low-resolution photographs. According to the article, super-resolution can be used for a variety of purposes, including improving the performance of existing medical imaging equipment and restoring historical family pictures.

Cascaded Diffusion Models (CDM)

After demonstrating the efficiency of the SR3 model, the Brain Team applied the model to the production of class-conditional images. “A class-conditional diffusion model trained on ImageNet data to generate high-resolution natural images,” the researchers say in their paper.

According to the blog post, Google created CDM as “a cascade of many diffusion models” since ImageNet was a tough dataset with a high degree of entropy. The model is a collection of various diffusion models that can be used to generate images with progressively higher levels of detail. It begins with a typical diffusion model at the lowest resolution and progresses through a series of super-resolution models that can incrementally upscale the image and add higher resolution details as the resolution of the image increases.

As part of their CDM implementation, Google employs a novel data augmentation approach known as “conditioning augmentation,” which is supposed to significantly improve the sample quality findings of the CDM test.

Google hopes to improve natural image synthesis, which has a wide range of uses but also creates design issues, with the introduction of these models. With SR3 and CDM, we have advanced the performance of diffusion models to the forefront of the field on benchmarks for super-resolution and class-conditional ImageNet production, the researchers stated in a blog post.


Leave a Reply

Your email address will not be published. Required fields are marked *