Google has shared on its blog the news about new algorithms based on artificial intelligence that can improve images by 16 times.
One of the algorithms was named SR3. It works on the principle of repeated refinement. So, the picture is loaded in low resolution, and the neural network, learning the methods of image distortion, adding noise to it, and then getting rid of them. The developers noted that the algorithm copes with its task better than existing generative algorithms, including PULSE and FSRGAN.
In addition to SR3, the company also has representatives of another diffusion model – CDM. She, in turn, was trained using millions of high-resolution pictures from the ImageNet database. This neural network works on the principle of a waterfall approach. So, the improvement of pictures occurs in several stages: the photo is improved from 32 × 32 to 64 × 64 and up to 256 × 256, or from 64 × 64 up to 1024 × 1024, that is, 16 times.
Read also:
Scientists have calculated the timing of the appearance of a pandemic that will destroy humanity
The best Realme watches are several times cheaper than the flagships of Samsung, Apple, etc. China won everyone, or is there a catch?
The largest application for watching pirated films on smartphones in Russia closed
The specialist explained when tea with lemon balm becomes dangerous for humans
Named 5 mushrooms that prolong youth and increase potency
