XPixelGroup/DiffBIR
DiffBIR
DiffBIR is a diffusion-based image restoration framework designed for blind image reconstruction. It utilizes generative diffusion priors to recover high-quality images from sources with unknown or complex degradations without requiring explicit degradation models.
The system includes specialized models for face restoration, enabling the recovery of facial landmarks, textures, and backgrounds in degraded portraits. To support high-resolution outputs on hardware with limited memory, it employs a tiled image upscaler that divides images into smaller patches during sampling.
The framework covers a multi-stage restoration pipeline and generative image upscaling. It incorporates capabilities for restoration model training and the application of specialized weights to optimize enhancement for specific scenes.
Features
- Blind Restoration Models - Provides a framework for restoring images without prior knowledge of the specific degradation applied.
- Diffusion Image Restoration Frameworks - Provides a complete diffusion-based framework for reconstructing high-quality images from degraded sources.
- Degradation Mappings - Implements techniques to map degraded visual inputs into clean latent spaces for blind image reconstruction.
- Generative Upscalers - Uses generative diffusion techniques to reconstruct missing high-resolution details during image upscaling.
- Image Resolution Reconstruction - Estimates original source pixels to restore fine details to degraded images using generative diffusion priors.
- Tiled Resolution Scaling - Uses tiled sampling to scale large images to high resolutions while preventing system memory exhaustion.
- Generative Priors - Utilizes pretrained generative networks as high-quality structural priors to guide image restoration.
- Tiled Upscaling - Increases image resolution using a tile-based refinement process to maintain spatial consistency and memory efficiency.
- Face Restoration - Provides generative restoration of facial structures to recover fine details in degraded portraits.
- Blind - Enhances facial features and recovers backgrounds in portraits without requiring prior face alignment.
- Diffusion Model Training - Includes workflows for training diffusion models to combine restoration with generative capabilities.
- Image Restoration Model Training - Provides capabilities to train multi-stage pipelines that integrate generative components for image restoration.
- Image Restoration Pipelines - Employs a multi-stage pipeline of specialized models to iteratively remove image artifacts and refine details.
- Large Scale Processing - Handles massive high-resolution images on limited hardware through distributed tiled processing.
- Face Restoration - Blind image restoration with generative diffusion priors.
- Image Restoration - Blind image restoration with generative diffusion priors.
- Inverse Problems - Listed in the “Inverse Problems” section of the Awesome Diffusion Models awesome list.
