Google has unveiled a new AI-powered image upscaling technique that improves the clarity of low-resolution images.
Brain Team researchers released two diffusion models to generate high-fidelity images in a post on Google’s AI blog titled “High Fidelity Image Generation Using Diffusion Model.”
The research team provided two linked approaches that push the image synthesis quality boundaries for diffusion models. Super-Resolution through Repeated Refinements (SR3) and Cascaded Diffusion Models are the two models (CDM).
The first is the SR3, a super-resolution diffusion model that takes a low-resolution image as input and generates a comparable high-resolution image from pure noise. Super-resolution offers a wide range of applications, from recovering old family pictures to upgrading medical imaging systems.
The model was trained on an image corruption process in which noise is gradually added to a high-resolution image until only pure noise remains, according to the research team. It then learns to reverse this process, starting with pure noise and gradually eliminating noise to obtain a goal distribution using the supplied low-resolution image as guidance.
SR3 produces strong benchmark performance on the super-resolution challenge for face and natural images when scaling to resolutions 4x–8x that of the input low-resolution image with large-scale training.
Meanwhile, Google employed these SR3 models for class-conditional picture generation after observing their effectiveness with SR3.
CDM is a class-conditional diffusion model that has been trained on ImageNet data to generate high-resolution natural images. “Because ImageNet is a tough, high-entropy dataset, we developed CDM as a cascade of many diffusion models,” Google explained in a blog post.
Furthermore, this cascade approach involves chaining together multiple generative models across multiple spatial resolutions: one diffusion model that generates data at a low resolution, followed by a series of SR3 super-resolution diffusion models that gradually increase the resolution of the generated image to the highest resolution.
“We have pushed the performance of diffusion models to state-of-the-art on super-resolution and class-conditional ImageNet generation benchmarks using SR3 and CDM,” Google added.
Google will continue to push the boundaries of diffusion models for a wide range of generative modeling issues.
Source: Google AI Blog via PetaPixel