SUPIR Extends to Excellence: Practical Model Scaling for Restoring Photorealistic Images in the Wild

I. Tutorial Introduction

SUPIR (Scaling-Up Image Restoration) is a groundbreaking method for image restoration and quality enhancement. Its constructed models leverage large-scale generative model Stable Diffusion XL (SDXL) along with model scaling techniques, utilizing deep learning and multimodal approaches to achieve high-quality recovery of low-resolution images. It can upscale low-resolution images to high resolution while preserving details and realism. SUPIR handles various complex degradation scenarios such as blur, noise, and compression, thereby enabling high-fidelity image reconstruction by upscaling low-resolution inputs into detailed, realistic high-resolution outputs.

The method also supports fine-grained control over image restoration via textual prompts, allowing adjustments based on user input. SUPIR was jointly developed by researchers from the Shenzhen Institute of Advanced Technology at the Chinese Academy of Sciences, Shanghai AI Laboratory, University of Sydney, Hong Kong Polytechnic University, Tencent ARC Lab, and The Chinese University of Hong Kong. The related paper titled "[Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild]" (https://arxiv.org/abs/2401.13627)) has been accepted by CVPR 2024.

This tutorial uses dual RTX 5090 GPUs. The maximum allowed file size is 1.5 MB.

👉 The project provides two model variants:

• SUPIR-v0Q: Default training settings. Offers higher generalization capability and better image quality under most circumstances.

• SUPIR-v0F: Trained using degraded lighting conditions. Stage 1 encoder SUPIR-v0F retains more detail even when illumination quality degrades.

Project Examples

II. Running Steps

1. After starting the container, click the API address to access the web interface.

If "Models" does not appear, it indicates that the model is initializing due to its large size. Please wait approximately 5 minutes before refreshing the page.

2. Once inside the webpage, you can interact with the model through dialogue.

Usage Instructions

Citation Information

Citation information for this project is provided below:

@misc{yu2024scaling,
  title={Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild}, 
  author={Fanghua Yu and Jinjin Gu and Zheyuan Li and Jinfan Hu and Xiangtao Kong and Xintao Wang and Jingwen He and Yu Qiao and Chao Dong},
  year={2024},
  eprint={2401.13627},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}