Image Super-Resolution via Iterative Refinement (SR3)

Key Features

• Diffusion-Based Super-Resolution: Utilizes denoising diffusion probabilistic models (DDPM) for image enhancement.
• Iterative Refinement: Employs a stochastic denoising process to progressively improve image quality.
• U-Net Architecture: Leverages a U-Net model trained on denoising tasks at varying noise levels.
• High Magnification Factors: Supports super-resolution tasks at 4×, 8×, and 16× magnification.
• High-Fidelity Outputs: Achieves photo-realistic results, with human evaluations indicating a fool rate close to 50% on 8× face super-resolution tasks.
• PyTorch Implementation: Provides an unofficial implementation in PyTorch for ease of use and customization.