If artificial intelligence is improving, it is image generation. Google has released AI-based image application technology that is said to enhance the quality of low resolution images. Simply put, Google’s AI works on a new technology that completely translates pixelated photos into high-resolution photos. In a post on Google’s AI blog, researchers on the Brain team introduced two diffusion models for creating high-resolution image images. The two models are Image Super-Resolution (SR3) and Cascade Diffusion Model (CDM).
The SR3 model, for example, is trained to transform a low-resolution image into a high-resolution image that outperforms current in-depth generative models such as Generative Rivals (GAN) in human evaluation.
Researchers at Google’s Brain Team have published a post on Google’s AI blog describing the SR3 and CDM diffusion models. The SR3 said to be a super-resolution diffusion model that takes low resolution image input and produces an ideal high resolution image from pure sound. The model is trained in the process of adding an image to a high resolution image until only pure sound is left. The SR3 model then reverses the process “starting with pure sound and gradually removing sound and reaching the target distribution under the guidance of the input low-resolution image.”
Google has shared some notable examples of scaling 64×64 pixel resolution image to 1,024×1,024 pixel resolution photo using SR3.The end result of the 1,024×1,024 pixel resolution output output, especially for the face and natural images, is quite impressive. The tech giant says the SR3 can achieve strong benchmark results in the super-resolution task for face and natural images when scaled to 4x to 8x higher resolutions.
The CDM Diffusion model is trained on image net data to create high resolution natural images. Because Image net is a complex and high entropy dataset, Google built CDM as a cascade of multiple diffusion models. This cascade approach involves combining multiple generative models with multiple spatial resolutions. The chain includes a diffusion model that generates data at a lower resolution, followed by a series of SR3 super-resolution diffusion models, and a diffusion model that gradually increases the resolution of the generated image to a higher resolution.
Google says Gaussian sound and Gaussian blur apply to the low-resolution input image of each super-resolution model in the cascading pipeline. This process is called conditioning augmentation, which enables better and higher resolution sample quality for CDMs.
Google says SR3 and CDM has “advanced the performance of diffusion models in super-resolution and class-conditioned image net generation benchmarks.”
Read interesting news, reviews as well as tips & tricks on TechnoBugg website, and stay updated with the latest happenings of the tech world on the go with Technobugg App. Also follow on Google News and join our Telegram channel for the latest updates.