High Fidelity Image Generation Using Diffusion Models by Google
Google has published an AI-based image upscaling technology that enhances the quality of low-resolution images.
Google’s Brain Team researchers have introduced two diffusion models to generate high fidelity images. The two models are image super-resolution (SR3) and cascaded diffusion models (CDM).
Google has posted this article for public visibility.
What is Image Super-Resolution?
The Image Super-Resolution is a super-resolution diffusion model. It takes an input of a low-resolution image and builds a corresponding high-resolution image from pure noise. The machine uses a process of image corruption.
In image corruption, noise is consistently added to a high-resolution image until only pure noise remains. This process is then reversed by eliminating the noise. It reaches a target distribution through the guidance of the input low-resolution image. Also read Bank holidays in September 2021
Some of the results shown by Google’s research team are very impressive. It showcases how this method can be used to effectively improve the image quality of low-resolution images.
As per the post, the super-resolution can have multiple applications. This can include enhancing the existing medical imaging systems and restoring old family portraits.
Cascaded Diffusion Models (CDM)
Once the SR3 model had shown effectiveness, the Brain Team used the model for class-conditional image generation.
The Cascaded Diffusion Model is a class-conditional diffusion model. This CDM trains on ImageNet data to generate a high-resolution natural image.
Google built CDM as “a cascade of multiple diffusion models” since ImageNet was a difficult, high-entropy dataset. The model is a combination of multiple diffusion models that can generate images of increasing resolution.
CDM follows the standard diffusion model at the lowest resolution. This process is followed by a sequence of super-resolution models that can upscale the image and add higher resolution details.
Google uses a new data augmentation technique, called “conditioning augmentation”. This improves the sample quality results of CDM.
Now Google is looking to improve the natural image synthesis that has wide-ranging applications but poses design challenges. Yes, our parent’s old photographs now have a chance of getting a new life.