Home
AI
Featured

Low-resolution images are transformed into high-quality images using Google's new AI models.

2 minute read

Google-ai-turns-low-res-images-into-high-res-ones-feat.

We've seen numerous AI tools and methods to improve image technology as researchers push their limitations to build sophisticated artificial intelligence (AI) technologies. We've seen AI technologies that immediately de-blur pictures and eliminate backgrounds from images. Google has now built two AI-based tools that use diffusion models to transform low-resolution pictures into high-resolution shots.

The two new methods, dubbed Super-Resolution through Repeated Refinements (SR3) and Cascaded Diffusion Models (CDM), were recently created by Google Research's Brain Team. The Mountain View behemoth recently detailed both technologies in an in-depth blog post on its AI forum. It's comparable to the prior AI program that Duke University of North Carolina academics developed earlier this year.

Starting with the SR3 model, it's essentially a super-resolution diffusion model that can turn low-resolution pictures into high-resolution ones using only noise. It takes a low-resolution image as input and gradually adds noise to it until only pure noise remains, utilizing an image corruption method that it was trained with. The procedure is then reversed, and the noise is removed to obtain the goal image, using the low-res input image as a reference.

The business claims that it was able to obtain strong benchmark results on the super-resolution challenge for face and natural pictures after large-scale training of the SR3 model. A 64 × 64 input image may be converted to a 1024 x 1024 output image using the model. To explain the process, Google posted a short video below that shows the SR3 model in operation.



The Cascaded Diffusion Model (CDM), the second AI model, is a class-conditional diffusion model that was trained on ImageNet data. By chaining several generative models over different spatial resolutions, the model can generate high-resolution natural pictures.

The CDM model generates data at a low resolution using one diffusion model, followed by a series of SR3 super-resolution diffusion models. A low-res image's resolution is steadily increased until it reaches its maximum resolution. To have a better sense of the picture generating process, look at the GIF attached below.

Aside from the two models mentioned above, Google AI researchers created a new data augmentation approach known as conditioning augmentation. It uses Gaussian noise and Gaussian blur to improve the sample quality outcomes of CDM. It also keeps each super-resolution model from overfitting to its lower-resolution conditioning input. It improves the quality of CDM's high-resolution samples.

With the aforementioned AI-based image enhancement models, Google claims to have advanced the state-of-the-art of diffusion models on super-resolution and class-conditional ImageNet generation benchmarks. In the future, the researchers will put these models to the test for more generative modeling challenges.

Pathakji...

Post a Comment