SUPIR (Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild) is a research project focused on developing practical algorithms for high-quality, photo-realistic image restoration and upscaling. It targets researchers and practitioners in computer vision and image processing who need to achieve state-of-the-art results on diverse, real-world images.

How It Works

SUPIR employs a two-stage diffusion model architecture. The first stage acts as an encoder, capturing essential image details, while the second stage performs the restoration and upscaling. This approach leverages large pre-trained models like SDXL and LLaVA for enhanced understanding and generation, allowing for fine-grained control over the restoration process through adjustable parameters and prompts.

Quick Start & Requirements

• Installation: Clone the repository and install dependencies via pip install -r requirements.txt within a conda environment (Python 3.8 recommended).
• Prerequisites: Requires PyTorch, Hugging Face Transformers, and specific pre-trained models (SDXL, LLaVA CLIP, etc.). Automatic download is supported if Hugging Face paths are configured correctly.
• Resources: Inference can be memory-intensive, with options for reduced VRAM usage (--loading_half_params, --use_tile_vae, --load_8bit_llava).
• Links: Project Page, Online App

Highlighted Details

• Achieves photo-realistic image restoration and upscaling "in the wild."
• Offers two model variants: 'Q' (generalization) and 'F' (light degradation).
• Supports customizable prompts and parameters for fine-tuned control.
• Includes a Gradio demo for interactive use and an online application.

Maintenance & Community

The project is associated with multiple institutions including Shenzhen Institute of Advanced Technology and Shanghai AI Laboratory. Contact emails for inquiries are provided.

Licensing & Compatibility

The software is explicitly declared for Non-Commercial Use Only. Commercial use requires prior written permission. This condition is added to any applicable open-source licenses.

Limitations & Caveats

The primary limitation is the strict non-commercial use clause, which may restrict integration into commercial products. The project is research-oriented, and while inference examples are provided, extensive production deployment guidance is not detailed.