Google’s AI technology can convert poor quality photos into high-resolution images

Search engine company Google has introduced diffusion models based on artificial intelligence (AI), with the help of which the quality of low-resolution images can be improved.
Two new diffusion models – Image Super-Resolution (SR3) and Cascaded Diffusion Models (CDM) – take the help of AI to produce high quality photos.
These models can be used to improve medical imaging systems, image classification and segmentation, from improving old family portraits.

Research team shared blog

A post has been shared on Google’s AI blog by researchers from the Google research team, which describes the SR3 and CDM diffusion models.
SR3 is called the super-resolution diffusion model, which allows low-resolution images to be converted into high-resolution images with existing data.
To do this, the model takes the help of the existing image data and makes necessary changes in them.

This is how the SR3 diffusion model works

Low-quality photos can be converted into high-quality images.

The SR3 model is trained with the Image Corruption Process, which adds noise to high-resolution images until only noise remains.
The SR3 model reverses this whole process and can produce high-quality photos with noise.
Actually, to resize an image, noise is included in it and the pixels become bigger, which affects the image quality.

This is how the second diffusion model has been prepared

The CDM Diffusion Model is trained with the help of ImageNet data, so that it can produce high-resolution natural images.
ImageNet is a difficult and high-entropy dataset, Google designed CDM as a cascade of multiple diffusion models.
With this cascade multiple generative models for different resolutions are created and combined together.
The models in the chain produce high-quality images that are applied to Gaussian Noise and Gaussian Blur to produce the final output.

Google shared results

The search engine company has also shared some examples related to this model, in which a photo of 64×64 pixels resolution was scaled with SR3 to convert it to 1,024×1024 pixels resolution photo.
The final 1,024×1024 pixels resolution output obtained with the help of this model is excellent and the details of the face visible in it are clearly visible.
The company says that the model can scale up to 4x and 8x high-resolutions of face and neural images.

AI Tech can become a part of future products

New technology related to Google’s AI can be made part of future products.
The use of AI has been seen before in the company’s services like Google Lens and Maps.
Image diffusion models can be used in areas ranging from safety to health, where a better image can be important and help.
However, Google has not said anything about this yet and is improving the existing tech.


Please enter your comment!
Please enter your name here