Google has posted on its blog a study by the internal Brain Team entitled "Creating High-Fidelity Imaging Using Diffusion Models." In this article, the researchers talk about the new advances they have made in scaling digital images without losing quality.
The Google Brain Team trained a machine learning model to turn low-resolution photos into detailed, high-resolution images with virtually no loss in quality. Experts believe that their development can be used for a variety of purposes, from improving old family photos to improving the quality of medical images.
The concept of diffusion models has been studied by Google since 2015, but until recently, the search giant preferred another family of AI training methods - deep generative models. The company found that the results of the new approach were markedly superior to existing technologies.
The new approach was designated SR3. Google says SR3 is an ultra-high-resolution diffusion model that creates a high-resolution image from pure noise based on the original low-resolution image. The model is trained in the process of image distortion, in which noise is gradually added to the image until only pure noise remains. The algorithm then reverses the process, gradually removing noise from the image, guided by the original low-resolution picture.
The SR3 was found to perform best when scaling portraits and nature shots. The algorithm allows you to achieve photorealistic images while increasing the resolution of portraits up to sixteen times.
Once Google was convinced how effective SR3 was, the company went even further with another approach called CDM, which is a conditionally class diffusion model. CDM is trained on data from ImageNet, which contains over 14 million high-resolution images. CDM proposes a cascading approach that first generates a low-resolution image, followed by SR3's work to create high-resolution images that are gradually increased to the highest possible. According to Google, an image with a resolution of 32 × 32 pixels can be enlarged to 256 × 256 pixels without noticeable loss, eight times. A picture with a resolution of 64 × 64 pixels was even scaled up to a resolution of 1024 × 1024 pixels, 16 times.
The results of AI work are really impressive. The final images, despite minor flaws, look really good and are perceived by most users as the original images.
0 Comments:
Post a Comment
Your comment and facebook share will be appreciated