synthetic degradation modeling

Real-world super-resolution (SR) faces two key challenges: (i) modeling complex, unknown degradations—such as spatially varying blur, heterogeneous noise, and color gamut shifts—using only unpaired low-resolution (LR) images; and (ii) the domain gap between synthetically generated and real-world degradations. Method: We propose TripleGAN, a fully unsupervised framework comprising three cascaded GANs: FirstGAN calibrates blur priors; SecondGAN performs domain-specific degradation transfer; and ThirdGAN reconstructs high-resolution (HR) images using pseudo-real, degradation-aligned training data. Contribution/Results: To our knowledge, this is the first method to jointly model implicit degradations and synthesize high-fidelity, degradation-matched training data in a purely unsupervised setting—without paired data or handcrafted priors. On RealSR and DRealSR benchmarks, TripleGAN achieves significant PSNR/SSIM gains over state-of-the-art methods. Reconstructed images exhibit sharp, natural textures with no oversmoothing artifacts, effectively bridging the synthetic-to-real degradation domain gap.

Addressing blind image restoration under unknown real-world degradations without ground-truth images, this work proposes a holistic solution. First, we introduce a learnable degradation chain estimator to accurately model complex, realistic degradations. Second, we design a consistency-driven, plug-and-play diffusion prior framework enabling end-to-end lightweight optimization. Third, we pioneer reference-free proxy metrics—MSE and LPIPS computed on synthetically degraded samples—that overcome the longstanding challenge of unreliable performance evaluation in blind restoration. To our knowledge, this is the first work unifying degradation modeling, restoration algorithm design, and no-reference assessment within a single coherent pipeline. Extensive experiments demonstrate substantial improvements in ranking accuracy over SOTA methods across multiple blind restoration benchmarks, significantly enhancing algorithmic assessability, comparability, and practical applicability.

Real-world distribution shifts—such as weather and illumination variations—severely degrade the robustness of deep learning models. However, collecting diverse, real-world degraded data is prohibitively expensive, prompting widespread reliance on synthetic degradation; yet its fidelity in reflecting real-world degradation effects remains unclear. Method: We construct the largest cross-domain (real vs. synthetic) semantic segmentation corruption benchmark to date, built upon Cityscapes and other datasets using the CorruptIO toolkit. We systematically evaluate 12 corruption types across multiple models and metrics (mIoU, RankCorr). Contribution/Results: We discover, for the first time, a strong correlation (ρ = 0.89) between model performance under real and synthetic corruptions. We further propose a corruption-type-level correlation analysis framework to characterize the applicability boundaries of synthetic degradation. All evaluation code, protocols, and benchmarks are publicly released to advance standardized robustness assessment.

Real-world image super-resolution (SR) suffers from a generalization bottleneck due to the mismatch between synthetic training data and authentic degradations. Existing degradation simulation or learning methods struggle to simultaneously achieve scalability, realism, and diversity. To address this, we propose RealDGen, an unsupervised framework built upon diffusion models that introduces a novel content-degradation disentanglement architecture. RealDGen generates large-scale, realistic, and diverse paired LR-HR training data using only unpaired real-world low- and high-resolution images—requiring neither paired supervision nor explicit degradation priors. It jointly integrates content extraction, prior-free degradation modeling, and reconstruction. Evaluated on multiple real-world SR benchmarks, models trained on RealDGen-synthesized data consistently outperform those trained on conventional datasets. The generated degradations comprehensively cover mixed noise, blur, and compression artifacts, achieving state-of-the-art fidelity and diversity in both visual quality and degradation coverage.

This paper addresses the poor generalization of learning-based image restoration methods in real-world scenarios—a limitation stemming from significant domain shift between synthetic training data and real images. To bridge this gap, we propose a novel domain adaptation paradigm tailored to the noise space of diffusion models. Our key contributions are: (1) the first “denoising-as-adaptation” mechanism, which progressively aligns restoration outputs of synthetic and real images toward the clean distribution via multi-step conditional denoising and a domain-aligned diffusion loss; and (2) a channel-shuffling layer coupled with residual-swap contrastive learning to implicitly blur domain boundaries and suppress shortcut feature dependencies. Evaluated on denoising, deblurring, and deraining tasks, our method substantially outperforms existing domain-adaptive and blind restoration approaches, achieving state-of-the-art generalization performance on real-world images.

Existing image restoration models struggle to generalize to real-world complex degradation scenarios due to limitations in the scale and distribution of training data. To address this, this work presents the first large-scale training dataset encompassing nine categories of real-world degradations and leverages a large-scale image editing model architecture trained with explicit modeling of real degradations alongside strategies to preserve content consistency. Furthermore, we introduce RealIR-Bench, a new evaluation benchmark that jointly assesses degradation removal efficacy and content fidelity. Experimental results demonstrate that the proposed method significantly outperforms existing open-source models on RealIR-Bench, achieving state-of-the-art performance among open approaches and substantially narrowing the gap with closed-source counterparts.

This work addresses the performance bottleneck of vision models caused by the scarcity of high-quality annotated data by proposing a modular synthetic data engine. The approach integrates controllable diffusion-based generation, multi-stage filtering, uncertainty-driven sample selection, and human-in-the-loop verification to establish a “synthetic-first” strategy calibrated by real data. Implemented via a command-line interface, the pipeline ensures reproducibility and flexible deployment. Evaluated on human pose estimation, the method achieves significantly better performance than real-data-only baselines when using only minimal real annotations augmented with synthetic data. Although purely synthetic training still lags behind, the results validate the effectiveness and practical utility of the proposed data curation framework in low-resource scenarios.

This work addresses the critical gap in understanding whether synthetic images are truly interchangeable with real ones in model training and the absence of systematic evaluation frameworks to ensure their safe and effective use. The study systematically quantifies discrepancies between synthetic and real images across three dimensions: high-dimensional feature distributions, low-level statistical properties in color space, and model training dynamics. Building on these insights, the authors propose a pre-evaluation metric for synthetic data of unknown quality and a safety-aware data fusion strategy for training. Experiments demonstrate that carefully calibrated mixing ratios and integration methods of synthetic and real data can substantially enhance model performance and robustness, thereby offering both theoretical grounding and practical guidance for the reliable deployment of synthetic data in machine learning pipelines.

This work addresses the scarcity of high-quality paired data for real-world image restoration, a challenge exacerbated by the inability of existing synthetic data to accurately model complex real-world degradations. To overcome this limitation, the authors propose leveraging multimodal foundation models—such as Nano-Banana-2—combined with a vision-language-model-driven adaptive prompting strategy to generate high-fidelity, content-faithful “Generative Ground Truth” (GGT) from real low-quality images. Through a multi-stage quality control pipeline, they construct GGT-100K, a large-scale real paired dataset comprising 103,707 training pairs and 500 test pairs. Experimental results demonstrate that models trained or fine-tuned on GGT-100K exhibit significantly improved generalization in real-world scenarios, with particularly pronounced gains for generative restoration architectures.

Existing image deblurring methods struggle to balance realism and flexibility due to the limited photorealism of synthetic data or the complexity of real-world data acquisition. This work proposes a flexible and efficient framework for capturing realistic deblurring data by using handheld cameras to acquire blurry images, while a gimbal-mounted camera densely captures sharp images to reconstruct the 3D scene. Paired sharp images are then rendered using estimated camera poses. A key innovation is the introduction of a Blur-aware Pose Refinement (BPR) module, which significantly improves geometric alignment between blurry and sharp image pairs. Leveraging this framework, the authors construct RealDeblur—a high-quality, diverse dataset—and demonstrate that models trained on it substantially outperform state-of-the-art methods across multiple real-world benchmarks, exhibiting strong generalization capability.