As researchers push the boundaries of developing advanced artificial intelligence (AI) technology, they have seen several AI tools and systems to improve imaging technology. We’ve seen AI tools that instantly remove backgrounds from images and remove photo blur. Currently, Google has developed two AI-based tools based on a diffusion model that can convert low-resolution images into high-quality photos.
These two new technologies, called super-resolution with iterative purification (SR3) and cascade diffusion model (CDM), were recently developed by the Brain Team at Google Research. The Mountain View giant recently posted a detailed blog post on the AI forum detailing both technologies. This is similar to a previous AI algorithm developed earlier this year by researchers at Duke University in North Carolina.
Now, starting with the SR3 model, it is essentially a super-resolution diffusion model, Convert low resolution images from pure noise to high resolution images.. It takes a low resolution image as input and uses a trained image corruption process to gradually add noise to the image until only pure noise remains. Then reverse the process to start denoising and reach the target image with reference to the low resolution input image.
According to the company, extensive training of the SR3 model has allowed it to achieve strong benchmark results in super-resolution tasks for facial and natural images.Model may Converts 64×64 input image to 1024×1024 image.. To show the process, Google shared a short video showing how the SR3 model works. You can see this just below.
Second, the second AI model, the Cascade Diffusion Model (CDM), is a class-conditional diffusion model trained with ImageNet data. This allows the model to create high resolution natural images by chaining multiple generative models to multiple spatial resolutions.
In this process The CDM model uses one diffusion model to generate data at low resolution, Followed by a series of SR3 super-resolution diffusion models. This will gradually increase the resolution of low resolution images to the highest resolution. You can check the GIF attached below to get a better idea of the image generation process.
In addition to the above two models, Google AI researchers Developed a new data expansion technology It is called conditioning enhancement. Further improve CDM sample quality results by using Gaussian noise and Gaussian blur. In addition, it prevents each super-resolution model from overfitting low-resolution adjustment inputs. This improves the quality of CDM high resolution samples.
Therefore, with the AI-based image improvement model described above, Google states that it has pushed the limits of the diffusion model to the forefront of super-resolution and class-conditional ImageNet generation benchmarks. Researchers will further test the limitations of these models for more generative modeling problems.
