Researchers at Google’s Brain Team have come up with a breakthrough AI machine-learning technology that can upscale low quality images effectively and turn low-resolution photos to high-res using diffusion models. The applications of these technologies can range from restoring old family photos to improving medical imaging systems.
The first approach, called Super-Resolution via Repeated Refinement (SR3), takes a low-resolution image as input, and builds a corresponding high resolution image from pure noise using a stochastic denoising process.
The second approach, called Cascaded Diffusion Models (CDM), uses SR3 models for class-conditional image generation and is trained on ImageNet data to generate high fidelity images.
Google claims that these diffusion models perform better than deep generative models such as GANs, VAEs, and autoregressive models because they offer more training stability and promising sample quality results. Check out some of the before-after images below.
Model 1: Super Resolution (SR3)
Google AI Turns Low-Res Images To High-Res - Super Resolution SR3 Model (1)
Google AI Turns Low-Res Images To High-Res - Super Resolution SR3 Model (2)
Google AI Turns Low-Res Images To High-Res - Super Resolution SR3 Model (3)
Google AI Turns Low-Res Images To High-Res - Super Resolution SR3 Model (4)
Google AI Turns Low-Res Images To High-Res - Super Resolution SR3 Model (5)
Watch the AI in action
Model 2: Cascaded Diffusion Model (CDM)
Learn more about the AI:
- Google AI Blog post • Super Resolution Model (SR3) on GitHub • Cascaded Diffusion Model (CDM) on GitHub • SR3 Research Paper (PDF) • CDM Research Paper (PDF)
Share this post with a designer/photographer friend and voice your views in the comments below. All images and videos via Google AI Blog.
