This new Google method transforms pixelated photos into spectacular high-resolution images

If one thing is getting better at artificial intelligence, it is image generation. Based on an initial model, current artificial intelligences can generate a similar image, color OR scale to higher resolution. The latter is something that Google’s AI does very well, so much so that convert fully pixelated photos to high resolution photos.

A recent research from Google’s artificial intelligence department shows how new advances in this aspect allow creating incredible images. The company’s machine learning model is capable of take a photo with hardly any resolution and scale it to get unique details.

Diffusion models

At the time of scale photos by artificial intelligence there are different methods to achieve this. The one used by Google is one called broadcast models. It is a generative model that began to be implemented in 2015 but it was only recently that this approach has been useful for Google.

As they explain, the system takes as input a low resolution image and from there builds a high resolution image on its own. For this Google says that first they have trained the AI ​​to lower the resolution of the images and make them extremely pixelated. From there, “learn to reverse this process, starting with pure noise and progressively removing noise to achieve a target distribution through the low-resolution input image guide.”

With this method Google manages to especially improve the portraits of people. However, it goes a step further and with a second AI it is able to scale the photographs even more. Resolutions of 32 x 32 px are capable of reaching up to 1024 x 1024 px. For this, first transform them to 64 x 64 px and take that new photograph as a reference to upload to 128 x 128 px, then the process again and so on until the desired resolution.

The results are undoubtedly spectacular, it allows create genuinely detailed photos from virtually nothing. Although there are some minimal errors (for example when generating the transparencies of the glasses), the photographs can pass as real without any problem. In fact, if the context is not known, an ordinary person would probably not identify that they are scaled by an AI.

The use of this? Indeed, improving photographs taken by users is a straightforward and clear application for this. For example, to improve the resolution of photos taken with mobile cameras, which are not always the best. However also can be useful in other sectors such as medicine to improve medical photographs.

Via | Business Insider
More information | Google AI Blog