Google developers presented a new technology that can significantly improve the quality of the original image. The AI reproduces the missing portions of even the most compressed image pixel by pixel.
The development is carried out by the Brain Team. They presented two algorithms for obtaining a photo. The first is SR3, it is able to scale the image by re-refinement. The increase in resolution is due to the borrowing of the missing fragments from the Gaussian noise. AI training is based on image distortion techniques and the subsequent reverse procedure.
Another diffuse technology is CDM. In it, artificial intelligence was trained using millions of high-resolution images from the ImageNet database. Scaling here is done in stages. So, the original 32x32 pixel image is improved to 64x64, and then up to 256x256 (8 times). The original 64 × 64 can be enlarged in this way up to 1024 × 1024, that is, 16 times.
According to the developers, the new technology shows better results in comparison with modern methods of scaling using AI BigGAN-deep and VQ-VAE-2.
The team did not provide more detailed information about the presented technology, except for demo animations. The release date of the commercial version of the development is unknown.
The development is carried out by the Brain Team. They presented two algorithms for obtaining a photo. The first is SR3, it is able to scale the image by re-refinement. The increase in resolution is due to the borrowing of the missing fragments from the Gaussian noise. AI training is based on image distortion techniques and the subsequent reverse procedure.
Another diffuse technology is CDM. In it, artificial intelligence was trained using millions of high-resolution images from the ImageNet database. Scaling here is done in stages. So, the original 32x32 pixel image is improved to 64x64, and then up to 256x256 (8 times). The original 64 × 64 can be enlarged in this way up to 1024 × 1024, that is, 16 times.
According to the developers, the new technology shows better results in comparison with modern methods of scaling using AI BigGAN-deep and VQ-VAE-2.
The team did not provide more detailed information about the presented technology, except for demo animations. The release date of the commercial version of the development is unknown.
Login or register to post comments
Comments 0