Towards Realistic Data Generation for Real-World Super-Resolution
Long Peng, Wenbo Li, Renjing Pei, Jingjing Ren, Jiaqi Xu, Yang Wang,, Yang Cao, Zheng-Jun Zha
TL;DR
This paper introduces RealDGen, an unsupervised framework that generates realistic paired low- and high-resolution images, improving super-resolution models' performance in real-world scenarios by better mimicking real degradations.
Contribution
The paper presents a novel content-degradation decoupled diffusion model for realistic data generation in super-resolution, addressing limitations of previous simulation and learning-based methods.
Findings
RealDGen produces high-quality, realistic paired data.
Enhanced super-resolution performance on real-world benchmarks.
Effective decoupling of content and degradation in data generation.
Abstract
Existing image super-resolution (SR) techniques often fail to generalize effectively in complex real-world settings due to the significant divergence between training data and practical scenarios. To address this challenge, previous efforts have either manually simulated intricate physical-based degradations or utilized learning-based techniques, yet these approaches remain inadequate for producing large-scale, realistic, and diverse data simultaneously. In this paper, we introduce a novel Realistic Decoupled Data Generator (RealDGen), an unsupervised learning data generation framework designed for real-world super-resolution. We meticulously develop content and degradation extraction strategies, which are integrated into a novel content-degradation decoupled diffusion model to create realistic low-resolution images from unpaired real LR and HR images. Extensive experiments demonstrate…
Peer Reviews
Decision·ICLR 2025 Poster
- The paper is well-structured, making the methodology and findings easy to understand. - The paper achieves competitive SR results under an unpaired setting, which is practical and advantageous in real-world applications.
- The paper argues that content and degradation can be decoupled by content and degradation extractor. However, it is unclear how the content extractor’s encoder is guaranteed to capture only pure content information. - Equation (5) utilizes X_{hr}, but no explanation or justification is provided for its use in the main manuscript. - The paper would benefit from a comparison of the total parameters of RealDGen with those of other methods, as this would provide additional context on scalabilit
1. The paper identifies the challenges in SR, particularly the gap between the synthetic degradations used in training data and the degradations in real-world images. By positioning the problem within the existing literature, the authors provide a strong motivation for their approach. 2. The core idea of the paper lies in decoupling content and degradation in real-world images to generate realistic data for SR, which is a novel approach for this task. The proposed approach, RealDGen, provides a
1. The paper does not clearly define Real-world LR. Through experiments, we can see that the work is more focused on the SR problem in real-world photography, but the authors do not clearly define or explain it in the paper. This makes Figure 2 (b) somewhat difficult to understand. 2. The overall presentation of this paper shoude be improved. The symbols in the formula do not correspond well to the figures. Eq.5 is not reflected in Figure 2. The colors used in Figure 1 (b) is somewhat confusion
1. The proposed RealDGen framework effectively separates content and degradation features using a diffusion model, facilitating the generation of realistic low-resolution images that more accurately replicate real-world degradation patterns. 2. By utilizing a well-designed contrastive learning approach for degradation extraction and a reconstruction-based method for content extraction, the proposed RealDGen framework effectively decouples degradation and content features, thereby enhancing data
1. What is the technical contribution of the proposed method RealDGen? It seems that the effectiveness of the proposed methods mainly comes from the powerful diffusion model, and a similar idea of separating the degradation and content features has been investigated by previous methods [A]. 2. During the training phase of DDPM, the authors finetune partial parameters of the extractor. However, it is unclear the motivation and the effect of finetuning partial parameters. Please provide more discu
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Image Processing Techniques · Advanced Vision and Imaging · Image Processing Techniques and Applications
MethodsDiffusion