Paper deep dive

QUSR: Quality-Aware and Uncertainty-Guided Image Super-Resolution Diffusion Model

Junjie Yin, Jiaju Li, Hanfa Xing

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 26

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%

Last extracted: 3/13/2026, 12:58:51 AM

Summary

QUSR is a novel image super-resolution framework that integrates a Quality-Aware Prior (QAP) generated by an MLLM and an Uncertainty-Guided Noise Generation (UNG) module. The QAP provides semantic and degradation-specific guidance, while the UNG module adaptively modulates noise injection based on regional reconstruction difficulty, enabling high-fidelity and photorealistic image restoration.

Entities (5)

QUSR · framework · 100%Qwen2.5-VL-7B-Instruct · mllm · 98%Quality-Aware Prior · module · 95%Stable Diffusion · diffusion-model · 95%Uncertainty-Guided Noise Generation · module · 95%

Relation Signals (4)

QUSR → integrates → Quality-Aware Prior

confidence 100% · QUSR, which integrates a Quality-Aware Prior (QAP)

QUSR → integrates → Uncertainty-Guided Noise Generation

confidence 100% · QUSR, which integrates... an Uncertainty-Guided Noise Generation (UNG) module.

QUSR → basedon → Stable Diffusion

confidence 95% · The core backbone network of QUSR is based on the UNet denoising model from Stable Diffusion

Quality-Aware Prior → leverages → Qwen2.5-VL-7B-Instruct

confidence 95% · the QAP leverages an advanced Multimodal Large Language Model (MLLM) to generate reliable quality descriptions

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Diffusion-based image super-resolution (ISR) has shown strong potential, but it still struggles in real-world scenarios where degradations are unknown and spatially non-uniform, often resulting in lost details or visual artifacts. To address this challenge, we propose a novel super-resolution diffusion model, QUSR, which integrates a Quality-Aware Prior (QAP) with an Uncertainty-Guided Noise Generation (UNG) module. The UNG module adaptively adjusts the noise injection intensity, applying stronger perturbations to high-uncertainty regions (e.g., edges and textures) to reconstruct complex details, while minimizing noise in low-uncertainty regions (e.g., flat areas) to preserve original information. Concurrently, the QAP leverages an advanced Multimodal Large Language Model (MLLM) to generate reliable quality descriptions, providing an effective and interpretable quality prior for the restoration process. Experimental results confirm that QUSR can produce high-fidelity and high-realism images in real-world scenarios. The source code is available at this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

25,244 characters extracted from source content.

Expand or collapse full text

QUSR: QUALITY-AWARE AND UNCERTAINTY-GUIDED IMAGE SUPER-RESOLUTION DIFFUSION MODEL Junjie Yin, Jiaju Li, Hanfa Xing † School of BeiDou Research Institute, South China Normal University, China ABSTRACT Diffusion-based image super-resolution (ISR) has shown strong po- tential, but it still struggles in real-world scenarios where degrada- tions are unknown and spatially non-uniform, often resulting in lost details or visual artifacts. To address this challenge, we propose a novel super-resolution diffusion model, QUSR, which integrates a Quality-Aware Prior (QAP) with an Uncertainty-Guided Noise Gen- eration (UNG) module. The UNG module adaptively adjusts the noise injection intensity, applying stronger perturbations to high- uncertainty regions (e.g., edges and textures) to reconstruct com- plex details, while minimizing noise in low-uncertainty regions (e.g., flat areas) to preserve original information. Concurrently, the QAP leverages an advanced Multimodal Large Language Model (MLLM) to generate reliable quality descriptions, providing an effective and interpretable quality prior for the restoration process. Experimen- tal results confirm that QUSR can produce high-fidelity and high- realism images in real-world scenarios. The source code is available at https://github.com/oTvTog/QUSR. Index Terms— Image super-resolution, Diffusion model, Quality-Aware Prior, Uncertainty 1. INTRODUCTION Image restoration is a critical task that has seen significant devel- opment due to its importance to downstream tasks [1–3]. Among these, image super-resolution (ISR) aims to reconstruct a clear, high- quality (HQ) image from its degraded, low-quality (LQ) counterpart. The core challenge of this problem lies in its ill-posed nature, par- ticularly in real-world scenarios where the degradation process is both unknown and complex. To address this issue, existing super- resolution methods have explored a variety of advanced network ar- chitectures and sophisticated degradation models. Although meth- ods based on Generative Adversarial Networks (GANs) [4, 5] have significantly advanced the field of SR, they still face a bottleneck in enhancing perceptual quality, especially in generating fine-grained textures, often introducing numerous visual artifacts. This is pri- marily because their optimization objectives tend to prioritize pixel- wise fidelity over visual realism, and models trained on synthetic data struggle to generalize to real-world images due to the inherent domain gap. With the recent rise of diffusion models in the image genera- tion domain, their progressive denoising mechanism has offered a new paradigm for real-world tasks such as image super-resolution and restoration [3, 6–8]. In particular, large-scale pre-trained Text- to-Image (T2I) models, exemplified by Stable Diffusion (SD), have † Corresponding author. This work was supported by the National Natural Science Foundation of China (General Program) under Grant 42271470. been successfully applied to a variety of downstream tasks owing to their remarkable semantic understanding and powerful genera- tive priors [9, 10]. To effectively leverage the potential of these pre- trained models, many studies have begun to employ techniques like ControlNet [11] to guide and control the generation process. This has demonstrated an exceptional capability for producing photoreal- istic and detail-rich images in applications like super-resolution and image restoration [3, 7, 12–15]. Current diffusion-based SR methods face key limitations. Mod- els like StableSR [7] and DiffBIR [3], which condition on the low- resolution image, struggle to extract effective semantic information when the input is highly degraded. In contrast, methods such as SeeSR [13] and PiSA-SR [14] use external models to generate text prompts, but these semantic descriptions overlook crucial degrada- tion information (e.g., blur, noise) necessary for accurate restoration. Although XPSR [15] employs a dual-prompting mechanism with MLLM to describe both content and degradation, its performance heavily depends on the accuracy of the MLLM’s judgments. To enhance super-resolution performance, and inspired by XPSR [15] and UPSR [16], we identify a central challenge in existing methods: the difficulty of reconciling high-level semantic guidance with low-level spatial fidelity when utilizing priors from the LQ image. A sole reliance on high-level text prompts tends to overlook inherent image degradation, while direct feature extrac- tion from the LQ image is often corrupted by noise and artifacts. To address the limitations of current approaches, specifically their insufficient understanding of global semantic content and quality attributes coupled with their inability to spatially adapt to local reconstruction difficulty, we propose an innovative dual-guidance framework named QUSR. The main contributions of this paper can be summarized as follows: • We propose a Quality-Aware Prior generated by a powerful MLLM. This prior provides a comprehensive textual descrip- tion that encapsulates both the semantic content and the spe- cific degradation attributes (e.g., clarity, noise) of the input image, thereby furnishing the model with holistic semantic guidance. • We design an uncertainty-guided noise generation mecha- nism within a single-step residual diffusion framework. This mechanism estimates the restoration difficulty of each image region, applying minimal noise to flat areas to preserve fi- delity while injecting stronger noise into complex textures to stimulate detail synthesis. To complement this, we introduce an uncertainty loss function that relaxes the reconstruction constraint on complex regions during training, enabling the model to focus on generating plausible details. • On two real-world datasets, our method demonstrates supe- rior performance by generating images with both high fidelity and photorealism. arXiv:2603.09125v1 [cs.CV] 10 Mar 2026 2. METHOD We propose QUSR, a novel image super-resolution framework based on a residual diffusion model, integrating adaptive uncertainty guid- ance with quality-aware prompting. The overall architecture is de- picted in Fig.1. 2.1. Framework of QUSR The core backbone network of QUSR is based on the UNet denoising model from Stable Diffusion and is optimized through parameter- efficient Low-Rank Adaptation (LoRA) fine-tuning. The overall pro- cess begins with the VAE encoderE , which maps the input LQ im- age x lq to a latent representation z lq = E(x lq ). Subsequently, the initial latent representation z lq is perturbed by adaptive noise gen- erated by the uncertainty estimation module, producing a guided la- tent representationz g . Then, the UNet network takes this perturbed, guided latent representation z g as its primary input and, guided by the quality-aware prior C q , predicts the added noise residual ε g : ε g = f(z g ,t,C q ),(1) where t represents the diffusion timestep (t = 1), and the condi- tional embeddings C q are integrated into each layer of the UNet via a cross-attention mechanism. The high-resolution latent representation z hq is then obtained by subtracting the predicted residual from the original latent repre- sentation: z hq = z lq −ε g .(2) Finally, the final HQ image x hq =D(z hq ) is generated through the VAE decoderD. 2.2. Quality-Aware Prior Leveraging multimodal pre-training and joint optimization over large-scale datasets, MLLMs exhibit exceptional capabilities in se- mantic understanding [18, 19]. Building on this foundation, we employ Qwen2.5-VL-7B-Instruct [20] from Qwen to extract quality priors from LQ images, thereby enhancing the model’s ability to perceive image quality and address quality issues in LQ images more effectively. We utilize the prompt instruction: ”Please de- scribe this low-resolution image, evaluating its quality based on clarity, color, noise, and lighting.” This approach generates de- scriptions that encompass key information, including the general content, overall quality, sharpness, noise level, and lighting, as demonstrated in Fig.1(a), where Qwen2.5-VL-7B-Instruct delivers quality descriptions aligned with human perception. These quality descriptions are processed by the CLIP text en- coder to generate quality embeddings, defined as C q = E clip (P q ), whereP q represents the quality prompt. These embeddings serve as conditional inputs to the cross-attention layers of the UNet network. The specific formula is as follows: F ′ m = softmax Q(F m )· K(C q ) T √ L · V (C q ).(3) where Q(·), K(·), and V (·) denote the query, key, and value projec- tions, respectively. F m denotes the feature map at the m-th layer of the UNet network, and F ′ m represents the embedded output feature. 2.3. Uncertainty-Guided Noise Generation Inspired by [16, 21], this module adaptively generates and injects noise based on the estimated uncertainty of the input image. The process consists of two main stages: Uncertainty Map Generation and Adaptive Noise Formulation. 2.3.1. Uncertainty Map Generation The Uncertainty Estimation Module (UEM) utilizes a lightweight encoder-decoder architecture. The encoder comprises three 3× 3 convolutional layers, each followed by an Exponential Linear Unit (ELU), while the decoder mirrors this structure but omits the final activation function. The LQ image x lq is processed by the UEM to produce an initial, raw uncertainty mapU . U =D uem (E uem (x lq )),(4) The uncertainty mapU represents the pixel-wise aleatoric error- scale, underpinning both the spatially variant noise schedule and the heteroscedastic reconstruction objective. 2.3.2. Adaptive Noise Formulation To refine the uncertainty representation, the initial mapU is first projected into the latent space by the main VAE encoderE and scaled by a factor k, yielding the latent uncertaintyU l = k · E(U). A minimum noise constraint m ∈ [0, 1] is then applied to produce the final uncertaintyU f : U f = m + (1− m)·U l ,(5) This constraint prevents unstable perturbations that could arise from excessively low noise. Based on this final uncertainty, the noise stan- dard deviation σ ε is computed. To ensure numerical stability, a small constant δ > 0 is added: σ ε = p |U f | + δ,(6) This mechanism ensures that high-uncertainty regions (e.g., edges, textures) receive stronger noise perturbations to facilitate the recon- struction of complex details, while low-uncertainty regions (e.g., flat areas) receive minimal noise, thus preserving information from the original input. Finally, the guided latent representation z g is formed by adding scaled Gaussian noise to the initial latent representation z lq : z g = z lq +ε· σ ε · p.(7) where ε ∼ N(0, 1) is a standard Gaussian noise tensor and p is a scalar representing the perturbation strength. This entire process optimizes the trade-off between information preservation and detail reconstruction during super-resolution. 2.4. Loss Function To guide the effective training of our model, we design a composite loss function L. This function is composed of four weighted loss terms, designed to synergistically optimize the quality of the gener- ated images across multiple dimensions. The total objective function is defined as: L = λ 1 L 2 + λ 2 L lpips + λ 3 L csd + λ 4 L un ,(8) CLIP text Encoder LQ Images VAE Decoder VAE Encoder ...... ...a succulent plant with..., resulting in slight blurriness. The lighting is natural but uneven, causing... Noise is minimal, preserving the texture of the plant and wood... ...an elderly man with ... The image is low-resolution,... and loss of detail. There's also some blurriness, particularly around the edges of his facial features,... Please describe this low-resolution image, evaluating its quality based on clarity, color, noise, and lighting. VAE Encoder Quality Embeddings · + − Quality Embeddings Trainable Module Frozen Module Cross Attention Conv+ELU Conv ... ... UEM ... Fig. 1. Framework of QUSR. (a) Initially, a MLLM is employed to generate quality-aware priors, each describing the content and degradation of the corresponding LQ image. These priors are then processed by a CLIP text encoder to derive quality embeddings. (b) Subsequently,, the LQ image is processed by the QUSR framework. Its core is a single-step denoising UNet that, conditioned on the quality embeddings, predicts the noise residual (ε g ) from the guided latent representation (z g ). Furthermore, the UEM generates an uncertainty map to construct adaptive noise perturbations, guiding the model to focus more on complex regions of the image during restoration. where λ 1 ,λ 2 ,λ 3 , and λ 4 are the hyperparameters that weight the contribution of each loss term. The L2 loss (L 2 ) penalizes pixel- level differences to ensure content fidelity between generated and ground truth images. The LPIPS [22] loss (L lpips ) uses the LPIPS metric to enhance visual realism by assessing perceptual similarity in the deep feature space. The Classifier Score Distillation (CSD) [23] loss (L csd ) employs a pre-trained Stable Diffusion model as an im- plicit classifier, leveraging Classifier-Free Guidance (CFG) to ex- tract semantic gradients, ensuring super-resolved results align visu- ally and semantically with quality-aware prompts. Furthermore, we introduce an uncertainty loss (L un ) that lever- ages the estimated uncertainty information to guide the optimization process. This loss guides the model to maintain high reconstruc- tion fidelity in regions of low uncertainty (e.g., smooth backgrounds) while permitting some reconstruction error in regions of high uncer- tainty (e.g., complex textures). Its formula is defined as: L un =L 1 (x hq ·exp(−U n ),x gt ·exp(−U n ))+α·mean(U n ). (9) whereL 1 denotes the L1 loss function. The weight term exp(−U n ) is derived directly from the normalized uncertainty mapU n . This strategy prioritizes the reconstruction fidelity of low-uncertainty re- gions. The second term is a regularization term controlled by the hyperparameterα, which constrains the overall distribution of uncer- tainty to prevent the model from producing trivially high uncertainty across the entire image. 3. EXPERIMENTS Datasets and Metrics. Following SeeSR [13], we employ the LSDIR [24] dataset along with the initial 10k images from the FFHQ [25] dataset for training, with LQ-HQ training pairs of 512× 512 resolution pre-generated through the RealESRGAN [5] degradation pipeline. The test dataset comprises center-cropped LQ- HQ image pairs sourced from the RealSR [26] and DRealSR [27] datasets. All LQ images in the test sets are resized to 128× 128, while the corresponding HQ images are 512 × 512. Evaluation metrics include reference-based measures: PSNR, SSIM, LPIPS, DISTS, and FID, as well as no-reference measures: CLIPIQA, MUSIQ, and MANIQA. Implementation Details. The QUSR model is built upon the Sta- ble Diffusion 2.1 framework and is designed for ×4 image super- resolution tasks. Training is conducted on four 24GB NVIDIA RTX 3090 GPUs, using the Adam optimizer with an initial learning rate of 3× 10 −5 and a batch size of 4, totaling 15K iterations. The LoRA rank is set to 4. For the loss function, the weight coefficients are λ 1 = 0.5, λ 2 = 2, λ 3 = 2, and λ 4 = 0.3. Quantitative Analysis. In this section, we quantitatively evalu- ate our proposed QUSR against other leading diffusion-based meth- ods [7, 8, 13, 14, 17] on two real-world datasets: RealSR [26] and DRealSR [27]. To ensure a fair comparison, all competing methods are tested using their officially released pre-trained models, and all metric scores are sourced from PiSA-SR [14]. As shown in Table 1, our method demonstrates superior performance on both test bench- marks. Notably, on the DRealSR dataset, QUSR achieves state-of- the-art (SOTA) results across all metrics. For instance, compared to the second-best performing method, it reduces the FID score by 16.74 while increasing the MUSIQ score by 0.89, highlighting its superiority in generating images with both high fidelity and percep- tual quality. Visual Comparison. Fig.2. illustrates the visual comparison be- tween QUSR and other state-of-the-art super-resolution methods on real-world images. As can be seen, existing methods often struggle Table 1. Quantitative comparison among the state-of-the-art DM-based SR methods on real-world test datasets. The best and second-best results are highlighted in red and blue, respectively. The symbols↑ and↓ denote that higher and lower values are preferable, respectively. DatasetsMethodPSNR↑SSIM↑LPIPS↓FID↓CLIPIQA↑MUSIQ↑MANIQA↑ RealSR StableSR [7]24.690.70520.3091127.200.619565.420.6211 SeeSR [13]25.330.72730.2985125.660.659469.370.6439 SinSR [17]26.300.73540.3212137.050.620460.410.5389 OSEDiff [8]25.150.73410.2921123.500.669369.090.6339 PiSA-SR [14]25.500.74140.2672124.090.670270.150.6560 Ours25.540.72890.2974125.270.682469.170.6564 DrealSR StableSR [7]28.040.74600.3354147.030.617158.500.5602 SeeSR [13]28.260.76980.3197149.860.667264.840.6026 SinSR [17]28.410.74950.3741177.050.636755.340.4898 OSEDiff [8]27.920.78350.2968135.290.696364.650.5899 PiSA-SR [14]28.310.78040.2960130.610.697066.110.6156 Ours29.810.82000.2708113.870.708267.000.6415 StableSR LQHQ SeeSRSinSR OSE-DiffPiSA-SROurs StableSRSeeSRSinSR OSE-DiffPiSA-SROurs LQHQ Fig. 2. Visual comparison of our method (QUSR) with SOTA methods on the RealSR and DRealSR datasets. to generate accurate and realistic details, particularly in regions with complex edges and fine textures. In contrast, QUSR effectively miti- gates this issue, producing details that are both structurally more ac- curate and visually more natural. Furthermore, QUSR demonstrates superior fidelity when handling dense, repetitive textures and signif- icantly reduces visual artifacts. Ablation Study. We validate the efficacy of two core modules: the Quality-Aware Prior (QAP) and the Uncertainty-Guided Noise gen- eration (UNG). As presented in Table 2, we evaluate three model variants: the model without the QAP module (w/o QAP), without the UNG module (w/o UNG), and without both modules, which serves as our baseline. Table 2. Ablation study of Quality-Aware Prior and Uncertainty. MethodPSNR↑ SSIM↑ CLIPIQA↑ MUSIQ↑ MANIQA↑ w/o QAP30.190.82060.685366.630.6318 w/o UNG29.740.81790.690666.580.6392 Baseline29.050.80710.674565.540.6268 Ours29.810.82000.708267.000.6415 Removing the Quality-Aware Prior (w/o QAP) leads to a slight increase in fidelity-oriented metrics such as PSNR/SSIM, but causes a significant drop in perception-oriented no-reference metrics like MUSIQ. This result validates that the global semantic and degrada- tion priors provided by the QAP module are crucial for guiding the model to generate realistic details that align with human perception. Removing the Uncertainty-Guided Noise generation (w/o UNG), in turn, results in a comprehensive decline across all metrics. This clearly demonstrates that the adaptive noise injection strategy of the UNG module plays a decisive role in the fine-grained reconstruc- tion of complex textures and in preventing the image from being over-smoothed. 4. CONCLUSION This paper introduces QUSR, a diffusion-based framework for image super-resolution. We leverage Qwen2.5-VL to generate a Quality-Aware Prior, providing comprehensive descriptions of con- tent and degradation to offer effective global semantic guidance. Concurrently, an Uncertainty-Guided Noise module is proposed to adaptively modulate noise injection intensity, prioritizing the re- construction of complex textures while preserving flat regions. By integrating these two mechanisms, QUSR balances high-level se- mantic guidance with low-level spatial fidelity, achieving superior performance and photorealism in real-world tasks. 5. REFERENCES [1] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou, “Snowformer: Scale-aware transformer via context interaction for single image desnowing,” arXiv preprint arXiv:2208.09703, vol. 2, p. 3, 2022. [2] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Jingxia Jiang, Erkang Chen, and Peng Chen, “Msp-former: Multi-scale pro- jection transformer for single image desnowing,” in ICASSP. IEEE, 2023, p. 1–5. [3] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong, “Diff- bir: Toward blind image restoration with generative diffusion prior,” in ECCV. Springer, 2024, p. 430–448. [4] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy, “Esrgan: Enhanced super-resolution generative adversarial networks,” in ECCV workshops, 2018, p. 0–0. [5] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan, “Real-esrgan: Training real-world blind super-resolution with pure synthetic data,” in ICCV, 2021, p. 1905–1914. [6] Sixiang Chen, Tian Ye, Kai Zhang, Zhaohu Xing, Yunlong Lin, and Lei Zhu, “Teaching tailored to talent: Adverse weather restoration via prompt pool and depth-anything constraint,” in ECCV. Springer, 2025, p. 95–115. [7] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy, “Exploiting diffusion prior for real-world image super-resolution,” International Journal of Computer Vision, vol. 132, no. 12, p. 5929–5949, 2024. [8] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang, “One-step effective diffusion network for real-world image super-resolution,” NeurIPS, vol. 37, p. 92529–92553, 2024. [9] Yunlong Lin, Zixu Lin, Haoyu Chen, Panwang Pan, Chenxin Li, Sixiang Chen, Kairun Wen, Yeying Jin, Wenbo Li, and Xinghao Ding, “Jarvisir: Elevating autonomous driving per- ception with intelligent image restoration,” in CVPR, 2025, p. 22369–22380. [10] SiXiang Chen, Jianyu Lai, Jialin Gao, Tian Ye, Haoyu Chen, Hengyu Shi, Shitong Shao, Yunlong Lin, Song Fei, Zhaohu Xing, et al., “Postercraft: Rethinking high-quality aesthetic poster generation in a unified framework,” arXiv preprint arXiv:2506.10741, 2025. [11] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala, “Adding conditional control to text-to-image diffusion models,”in ICCV, 2023, p. 3836–3847. [12] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang, “Pixel-aware stable diffusion for realistic image super- resolution and personalized stylization,” in ECCV. Springer, 2024, p. 74–91. [13] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang, “Seesr: Towards semantics-aware real-world image super-resolution,”in CVPR, 2024, p. 25456–25467. [14] Lingchen Sun, Rongyuan Wu, Zhiyuan Ma, Shuaizheng Liu, Qiaosi Yi, and Lei Zhang, “Pixel-level and semantic-level ad- justable super-resolution: A dual-lora approach,” in CVPR, 2025, p. 2333–2343. [15] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou, “Xpsr: Cross-modal priors for diffusion-based image super-resolution,” in ECCV. Springer, 2024, p. 285–303. [16] Leheng Zhang, Weiyi You, Kexuan Shi, and Shuhang Gu, “Uncertainty-guided perturbation for image super-resolution diffusion model,” in CVPR, 2025, p. 17980–17989. [17] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lan- qing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen, “Sinsr: diffusion-based image super-resolution in a single step,” in CVPR, 2024, p. 25796–25805. [18] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al.,“Q-bench: A benchmark for general-purpose foundation models on low-level vision,” arXiv preprint arXiv:2309.14181, 2023. [19] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong, “Depicting beyond scores: Advancing image quality assessment through multi-modal language mod- els,” in ECCV. Springer, 2024, p. 259–276. [20] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al., “Qwen2. 5-vl technical report,” arXiv preprint arXiv:2502.13923, 2025. [21] Sixiang Chen, Tian Ye, Jinbin Bai, Erkang Chen, Jun Shi, and Lei Zhu, “Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks,” in ICCV, 2023, p. 13106–13117. [22] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang, “The unreasonable effectiveness of deep fea- tures as a perceptual metric,” in CVPR, 2018, p. 586–595. [23] Jonathan Ho and Tim Salimans, “Classifier-free diffusion guid- ance,” arXiv preprint arXiv:2207.12598, 2022. [24] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Deman- dolx, et al., “Lsdir: A large scale dataset for image restoration,” in CVPR, 2023, p. 1775–1787. [25] Tero Karras, Samuli Laine, and Timo Aila, “A style-based generator architecture for generative adversarial networks,” in CVPR, 2019, p. 4401–4410. [26] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang, “Toward real-world single image super-resolution: A new benchmark and a new model,” in ICCV, 2019, p. 3086– 3095. [27] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixi- ang Ye, Wangmeng Zuo, and Liang Lin, “Component divide- and-conquer for real-world image super-resolution,” in ECCV. Springer, 2020, p. 101–117.

Instant research discovery