Image Super-Resolution via Iterative Refinement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs.
• Existing methods often struggle to generate photorealistic details and high fidelity.
• Diffusion models have shown promise in generative tasks but require adaptation for image-to-image translation.

Purpose of the Study:

• To introduce SR3, a novel approach for image super-resolution using repeated refinement.
• To adapt denoising diffusion probabilistic models for the image-to-image translation task of super-resolution.
• To demonstrate SR3's effectiveness in generating high-resolution images with enhanced realism and detail.

Main Methods:

• SR3 employs a stochastic iterative denoising process, starting with Gaussian noise.
• A U-Net architecture is trained for denoising at various noise levels, conditioned on low-resolution input.
• The approach leverages denoising diffusion probabilistic models adapted for super-resolution.

Main Results:

• SR3 achieves strong performance on super-resolution tasks across different magnification factors for faces and natural images.
• Human evaluation on 8x face super-resolution shows SR3 achieving a near 50% fool rate, indicating photorealism.
• SR3 outperforms baseline methods in human evaluation and classification accuracy on a 4x ImageNet super-resolution task.

Conclusions:

• SR3 effectively generates photorealistic super-resolved images, surpassing current state-of-the-art methods.
• The approach demonstrates versatility, showing success in cascaded image generation tasks.
• SR3 represents a significant advancement in image super-resolution using diffusion models.