nachifur/RDDM
CVPR 2024: Residual Denoising Diffusion Models
What it solves
RDDM addresses the limitations of standard diffusion models for image restoration and generation. It provides a framework for improving the quality of restored images (such as removing raindrops, blur, or noise) and generating new images from scratch.
How it works
RDDM uses a residual denoising approach. It employs two UNet architectures to handle the "deraiding" (residual) and denoising processes. In some configurations, it can convert pre-trained DDIM (Denoising Diffusion Implicit Models) into RDDM to leverage existing models for image generation.
Who it’s for
Researchers and developers working in computer vision, specifically those focused on image restoration (deraining, deblurring, denoising) and generative AI for images.
Highlights
- Versatile Image Restoration: Supports multiple tasks including deraining (Raindrop dataset), deblurring (GoPro), and low-light enhancement (LOL).
- Multimodal Capabilities: Capable of both image generation (CelebA) and image restoration.
- DDIM Conversion: Ability to convert pre-trained DDIM models to RDDM via coefficient transformation.
- Broad Application Range: Includes experiments for image inpainting and image translation (e.g., dog-to-cat).