This repository provides PyTorch implementations for Generative Diffusion Prior (GDP) models, enabling unified image restoration and enhancement tasks. It targets researchers and practitioners in computer vision and deep learning, offering a flexible framework for tasks like super-resolution, deblurring, inpainting, colorization, low-light enhancement, and HDR recovery.

How It Works

The project leverages unconditional DDPMs pre-trained on ImageNet as generative priors. These diffusion models are guided by input degraded images to synthesize high-fidelity restored outputs. The approach allows for a unified framework across various restoration tasks by adapting the guidance mechanism and sampling process, demonstrating effectiveness for both linear and non-linear degradation problems.

Quick Start & Requirements

• Install: pip install -e .
• Prerequisites: PyTorch, Python 3.x. Requires downloading pre-trained DDPMs on ImageNet-256 (unconditional) and sample datasets (ImageNet-64, LSUN-Bedroom, LSUN-Cat).
• Links: Paper, Project Page

Highlighted Details

• Unified framework for diverse image restoration and enhancement tasks.
• Utilizes unconditional DDPMs as generative priors.
• Supports both linear and non-linear degradation models.
• Implemented based on OpenAI's Guided Diffusion repository.

Maintenance & Community

The project is associated with the CVPR 2023 paper "Generative Diffusion Prior for Unified Image Restoration and Enhancement." It is inspired by several other prominent diffusion model repositories.

Licensing & Compatibility

The repository's license is not explicitly stated in the README. Compatibility for commercial use or closed-source linking would require clarification of the licensing terms.

Limitations & Caveats

The README does not specify the exact license, which may impact commercial use. The implementation relies on specific pre-trained models and datasets that need to be downloaded separately.