Developers Brain Team told in the blog Google AI on the company’s achievements in the field of artificial intelligence. Experts have created a new image scaling technology that can reproduce high-quality photographs from low-resolution images.
For this, two diffusion models are used. The first is called SR3 or super-resolution by re-refinement. It takes a low-resolution image as input, and then gradually adds noise to it until the image takes on clear shapes. The model learns, reverses the process, starting from the beginning and gradually removing noise to achieve the stated result.
Google employees found the SR3 to perform well for upscaling portraits and nature photographs. At 8x zooming in on faces, the model performs better than generative algorithms such as PULSE and FSRGAN.
The second approach is called CDM or conditional class diffusion model. It reproduces high definition images using ImageNet data.
Google has published examples of work demonstrating cascading scaling of low-resolution images: 32 × 32 photos are upscaled to 64 × 64 and then up to 256 × 256, and 64 × 64 photos up to 256 × 256 and up to 1024 × 1024.
Read also: Google may develop its own ARM-based processors…
