title: [Paper Review] Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild
published: false
date: 2024-01-26 00:00:00 UTC
tags:
canonical_url: http://www.evanlin.com/paper-supir/
Paper Title: Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild
Scaling Up to Excellence: Amazing Image Restoration Paper
Proposes SUPIR (Scaling-UP Image Restoration), mainly with the following methods:
- Enhances restoration capabilities through prompts, even strengthening through Negative Prompts
- Over 20 million ultra-high-definition training materials.
The results show very good performance in terms of quality.
Paper: https://arxiv.org/abs/2401.13627
Website: https://supir.xpixel.group/
Quick Summary
- Research Topic: Achieving ultra-high-quality image restoration for real-world scenarios using generative priors and model scaling techniques.
- Dataset: Collected 20 million high-resolution, high-quality images, each with descriptive text annotations, for model training.
- Innovation: Proposes a method to guide image restoration through text prompts, expanding the application scope and potential of image restoration. Also introduces negative quality prompts to further improve perceptual quality. Also developed a restoration-guided sampling method to suppress fidelity issues encountered in generative restoration.
- Experimental Results: Demonstrates SUPIR's excellent image restoration effects and its novel ability to perform restoration operations through text prompts.
System Architecture
The main content of EDM Sampler with Restoration Guidance is as follows:
- EDM Sampler with Restoration Guidance is a sampling method for image restoration, based on the principle of diffusion models, and introduces a new hyperparameter τr to control the degree of guidance to low-quality images during the sampling process.
- Purpose: The purpose of this method is to improve the fidelity of the image while maintaining the perceptual quality of the image, even if the image does not deviate from the content of the low-quality image during the sampling process.
- Principle: The principle of this method is to mix the predicted result zt−1 with the low-quality image zLQ in a certain proportion in each step of sampling, according to the size of τr, so that zt−1 is close to zLQ while retaining the generated details and textures.
- Effect: The effect of this method has been verified through experiments on different synthetic and real low-quality images, and compared with other image restoration methods. The results show that this method achieves the best performance on non-reference metrics and also performs well on full-reference metrics. In addition, this method can also achieve flexible control of image restoration by adjusting the size of τr, thereby achieving different effects.
Top comments (0)