4KAgent: Agentic Any Image to 4K Super-Resolution

We sincerely thank Reviewer eAhT for highlighting the strengths of our work. We are glad that you found our agentic framework for 4K super-resolution systematic, well-structured, and practically relevant. Your recognition of our clear plan–execution–evaluation–reflection pipeline, the dedicated face restoration subsystem, and the flexibility we provide for user customization is greatly appreciated.

W1: Provide the fast/slow option to balance the performance and process time

A1: TL;DR: We develop a Fast4K mode for 4KAgent to balance the performance and process time.

Thank you for pointing this out. We developed a Fast4K mode based on 4KAgent. Specifically, when the image resolution is higher than the resolution threshold s set by the user (longer side > s), 4KAgent will disable long-running methods in the toolbox in order to reduce the processing time of 4KAgent. At the same time, this resolution threshold s can be set by the user, so the user can flexibly adjust it to balance performance and processing time.

Based on the Fast4K mode, we conduct an experiment on DIV4K-50 by setting different resolution threshold s (s = 1024 / 4096) in 4KAgent and evaluate the running time and performance. Experiment result is shown below (Running time is evaluated by RTX 4090 GPU):

It shows that, simply set s to 1024, the average running time of 4KAgent decreases from 1551.8s to 582.7s (62.3% ↓) with minimal effect on performance. Under this setting, 4KAgent is even faster than the model-based approach (DiffBIR).

W2: Unnatural region effect in edge for face restoration pipeline

A2: TL;DR: The face restoration pipeline will introduce boundary artifacts in a low frequency of occurrence, and the benefits of the face restoration pipeline in 4KAgent outweigh the impact of the boundary artifacts it introduces.

Thank you for pointing this out. When designing the face restoration pipeline in 4KAgent, we follow the ‘whole image restoration’ procedure from previous methods (e.g., GFPGAN [4], CodeFormer [5], DifFace [6]). Specifically, we use the toolbox in GFPGAN to extract faces in the image and paste them back after restoration. Therefore, it may introduce boundary artifacts at the edge. We conduct an experiment for this.

Specifically, we select 100 LQ images in the WebPhoto-Test dataset as test images. Then, we input these images to GFPGAN, CodeFormer, and DifFace for face restoration under the ‘whole image restoration’ procedure. We also collect result images from 4KAgent. After obtaining result images, we conduct a user study to explore the boundary artifacts. We define the boundary artifacts as: Jagged edges or color halos along the face boundary. We ask users to evaluate the result images and count the occurrence of the boundary artifacts. Then we calculate the average occurrence of the boundary artifacts. Experiment results are shown below:

It shows that SOTA face restoration methods and 4KAgent introduce boundary artifacts following the ‘whole image restoration’ procedure. However, the frequency of occurrence is low (less than 10%), and 4KAgent achieves the lowest frequency of occurrence with its system design.

We also conduct an ablation study of the face restoration pipeline. Specifically, we compare the result images from 4KAgent with different profiles (ExpSR-s4-P, ExpSRFR-s4-P) on the WebPhoto-Test dataset. The key difference between profiles and corresponding experiment results is shown below:

where ‘Face Restore’ controls the activation of the face restoration pipeline. It shows that when activating the face restoration pipeline, both the general image perceptual quality and face-related IQA metrics are enhanced. We also provide a visual comparison in Figure 8 in the Appendix, which further demonstrates the superiority of the face restoration pipeline by the enhanced facial and hair details from the image from 4KAgent by activating the face restoration pipeline.

Therefore, following the previous procedure, the face restoration pipeline in 4KAgent will introduce boundary artifacts with a low frequency of occurrence. However, the face restoration pipeline will enhance the overall image quality and face quality. We believe that the benefits of the face restoration pipeline in 4KAgent outweigh the impact of the boundary artifacts it introduces. We will further modify the face restoration pipeline to reduce the boundary artifacts it introduces.

W3: The role of the evaluation metrics should be further investigated

A3: Thank you for pointing this out. When designing the quality score used for evaluation in the 4KAgent system, we combined commonly used no-reference IQA metrics and HPSv2 scores (for human preferences). Experimental results show that 4KAgent achieves SOTA results, especially in terms of perceptual quality. In the future, we will further study the role of evaluation metrics, such as using block processing at ultra-high resolution to achieve more accurate evaluation results, and combining more advanced aesthetic scores.

W4: Provide examples of restoration plan and agenda for exhibition

A4: Thank you for pointing this out. We show some examples here.

For the first image in Figure 1 with tag ‘4K Upscaling’, its restoration plan and restoration sequence is: Denoising@SwinIR(σ=50) → Super-resolution@DiffBIR → Super-resolution@SwinIR (Real-ISR).

For the lower left image in Figure 1 with tag ‘Human Faces’, its restoration plan and restoration sequence is: Super-resolution@OSEDiff → Denoising@Restormer → JPEG Compression Artifact Removal@FBCNN (BQF) → Defocus Deblurring@DiffPlugin → Super-resolution@DiffBIR. BQF in FBCNN indicated 'Blind to Quality Factor'.

For the left image in Figure 4, its restoration plan and restoration sequence is: Super-resolution@PiSA-SR → Dehazing@RIDCP.

For the right image in Figure 4, its restoration plan and restoration sequence is: Deraining@Restormer → Denoising@SwinIR(σ=15) → Super-resolution → OSEDiff.

We will show more examples of image restoration plans and the intermediate images produced by each step in the Appendix in the next version of paper.

Q1: Observing the winning rate MoE to identify restoration models that suited for different domains

A5: Thank you for pointing this out. As we conduct experiments on images from different domains, we calculate the winning rate of SR methods from these experiments. The result is shown below:

It shows that when facing images from different domains, different SR methods win from the MoE policy. We believe that the results obtained by analyzing the winning rate can provide some insights into model selection for SR tasks in different fields of images.

Reference

[1] Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., ... & Dong, C. (2024). Diffbir: Toward blind image restoration with generative diffusion prior. ECCV.

[2] Wu, R., Sun, L., Ma, Z., & Zhang, L. (2024). One-step effective diffusion network for real-world image super-resolution. NeurIPS.

[3] Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., & Zhang, L. (2025). Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. CVPR.

[4] Wang, X., Li, Y., Zhang, H., & Shan, Y. (2021). Towards real-world blind face restoration with generative facial prior. CVPR.

[5] Zhou, S., Chan, K., Li, C., & Loy, C. C. (2022). Towards robust blind face restoration with codebook lookup transformer. NeurIPS.

[6] Yue, Z., & Loy, C. C. (2024). Difface: Blind face restoration with diffused error contraction. TPAMI.

We sincerely thank Reviewer wkHw for your positive feedback. We appreciate your recognition of the overall clarity of our writing, the interesting results presented, and the thorough evaluation of our proposed image enhancement system across a wide range of datasets. Your comments are encouraging and affirm the value of our work.

Q1 & L2: Follow the experimental settings of recent SOTA generative SR methods, and conduct a side-by-side comparison

A1: We thank the reviewer for pointing it out. In Section 3.2 in the paper, we adopt the same experimental setup as recent state‐of‐the‐art generative super-resolution (SR) models (e.g., SinSR [1], OSEDiff [2], PiSA‑SR [3]) and perform 4x SR on real-world image SR benchmarks: the RealSR and DRealSR datasets. Quantitatively, 4KAgent outperforms SOTA generative SR methods on perceptual metrics (NIQE, CLIPIQA, MUSIQ, and CLIPIQA). We also provide experimental results on the DRealSR dataset here:

Qualitative results, which we provided in section C.2 in the Appendix, further illustrate that 4KAgent produces images with sharper and more realistic details. This demonstrates that, despite the strong performance of existing generative SR approaches on real‐world images, 4KAgent surpasses them by decoupling the complex distortion in the image, performing the corresponding restoration task while using the Q-MoE strategy to select the image's highest quality, which yields superior perceptual quality.

For the 16x SR experiments in Section 3.4 and the 4K upscaling experiments in Section 3.5, 4KAgent operates in a training‑free manner: we apply it directly to the 16x SR task without additional model training. To ensure a fair comparison, we likewise treat the leading generative SR models as training‑free in this setting and evaluate them in both the 4x->4x and the 16x strategy. We presented visual comparison in the main paper, showing that 4KAgent produces more natural and realistic details. Quantitative results in sections C.5 and C.6 in the Appendix confirm that 4KAgent achieves the state of the art performance on NIQE, MUSIQ, and CLIPIQA metrics.

Q2: How natural language descriptions could benefit SR model for recovering low-level details?

A2: Natural language description is used in the reflection stage to select the image with the best quality (based on the alignment to the image description). It is not used in single SR model in 4KAgent system.

Q3: Provide ablation studies to see good results comes from good restoration plan or powerful tools, and analysis tool usage on real-world image super-resolution task

A3: Thank you for pointing this out. We conduct an ablation study on the DIV4K-50 dataset for 16x super-resolution. Specifically, we fix the plan as 'super-resolution → super-resolution' and apply different SOTA generative SR methods (DiffBIR [4], OSEDiff [2], PiSA-SR [3]) as the tool in each super-resolution step. We use 3 metrics for evaluation: NIQE, MUSIQ, and MANIQA-pipal. Experiment results are shown below:

It shows that using powerful tools can achieve good perceptual quality, but combining them with a good plan can further improve perceptual quality.

Q4: Report some information about runtime

A4: Thank you for pointing this out. As a multi-agent system, 4KAgent supports multi-GPU deployment. Specifically, 4KAgent assigns different agents (Perception Agent, Restoration Agent) to different GPUs to conserve memory. Most of our experiments were conducted using two NVIDIA RTX 4090 GPUs. Therefore, we evaluate the running time using NVIDIA RTX 4090 GPUs. The running time of 4KAgent depends on the chosen profile, the quality of the input image, and the length of the restoration plan. We report the fastest and slowest cases observed in our experiments. The fastest case is super-resolve images (x4) on the B100 dataset under ExpSR-s4-F profile. The slowest case is jointly restoring and upscaling LQ images on the DIV4K dataset to 4K resolution under Gen4K-P profile. The running time under these two cases is shown below:

For accelerating 4KAgent, we developed a Fast4K mode for 4KAgent. Specifically, when the image resolution is higher than the resolution threshold s set by the user (longer side > s), 4KAgent will disable long-running methods in the toolbox in order to reduce the processing time of 4KAgent. At the same time, this resolution threshold s can be set by the user, so the user can flexibly adjust it to balance performance and processing time.

Based on the Fast4K mode, we conduct an experiment on DIV4K-50 by setting different resolution threshold s (s = 256 / 1024 / 4096) in 4KAgent and evaluate the running time and performance. Experiment result is shown below:

It shows that, simply set s to 1024, the average running time of 4KAgent decreases from 1551.8s to 582.7s (62.3% ↓) with minimal effect on performance. Under this setting, 4KAgent is even faster than the model-based approach (DiffBIR).

Q5: Rain-streak removal should not automatically conducted in an image restoration system

A5: Thank you for pointing it out. One of the core contributions of 4KAgent is that it is a highly customizable system. By setting different profiles, 4KAgent can be customized to perform different image restoration tasks. We provide specific profile settings and examples in Section A.1 in the Appendix. Therefore, for images containing rain streaks, users can flexibly choose whether to perform rain removal.

L1: Handle real-world data

A6: Thank you for pointing this out. In section 3.4, we test 4KAgent on RealSRSet, where each low-quality image is collected from the real world. Therefore, it contains real-world noise and artifacts. We upscale these images by a scale factor of 16 and evaluate the result images by 3 metrics: NIQE, MUSIQ, and MANIQA-pipal. Experiment results are shown below:

It shows that 4KAgent performs on par or better than SOTA generative SR methods, indicating its ability to handle real-world data.

Reference

[1] Wang, Y., Yang, W., Chen, X., Wang, Y., Guo, L., Chau, L. P., ... & Wen, B. (2024). Sinsr: diffusion-based image super-resolution in a single step. CVPR.

[2] Wu, R., Sun, L., Ma, Z., & Zhang, L. (2024). One-step effective diffusion network for real-world image super-resolution. NeurIPS.

[3] Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., & Zhang, L. (2025). Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. CVPR.

[4] Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., ... & Dong, C. (2024). Diffbir: Toward blind image restoration with generative diffusion prior. ECCV.

[5] Zhu, K., Gu, J., You, Z., Qiao, Y., & Dong, C. (2024). An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809.

Dear Reviewer wkHw,

We are thankful for your detailed comments. Please find the response below.

Q1 & Q5: Although 4KAgent excels in no-reference IQA metrics, its LPIPS and DISTS still trail state-of-the-art generative SR methods, suggesting it hallucinates details rather than truly recovering them, and it lacks validation on real-world restoration benchmarks, undermining the fairness and credibility of the comparisons.

A1: TL;DR: 4KAgent offers two configurable modes: fidelity for maximizing traditional SR metrics (PSNR, SSIM) and perception for non-reference perceptual scores (NIQE, CLIPIQA, MUSIQ, MANIQA) while remaining competitive on LPIPS and DISTS. This flexible design addresses the fundamental perception–distortion trade-off, letting users choose higher accuracy or more creative “hallucinations” (e.g., enhancing AI-generated images and old photos). Across benchmarks (B100, RealSR, DRealSR, GenAIBench-4K, DiffusionDB-4K, WorldStrat), 4KAgent matches or exceeds state-of-the-art performance in the selected mode.

We thank the reviewer for the question. We’d like to clarify that 4KAgent is a highly customizable system (with a simple argparse-like config): when performing SR tasks, users can pre-set different modes: either prioritizing higher fidelity or higher perception quality, which corresponds to more faithful recovery or higher quality but arguably more creative (c.f. the reviewer’s point on the conditional generation), respectively. We’d also like to mention that it’s almost impossible to achieve the highest scores for both perceptual and fidelity metrics. This phenomenon has been established by the perception-distortion tradeoff papers [1]. We also observed in Table 1 in the main paper that the recent state-of-the-art diffusion-based models like OSEDiff and PiSA-SR can achieve high perceptual scores but still struggle on fidelity metrics. Because of this fundamental tradeoff, we have made this a configurable choice for the user: you can flexibly choose to be more accurate (i.e., the traditional SR problem), or more “creative” (we think this is very useful to handle AI-generated content and old photos where more “hallucination” is needed)

For the traditional SR problem, we follow the experiment setting from previous methods and present the experiment results on the B100 dataset here. The top two performances are marked in bold and italic. (Detailed experiment results on all traditional SR benchmarks are provided in Tables 5 ~ 9 in the Appendix.)

It shows that under the fidelity mode (4KAgent (F), choose to be more accurate), 4KAgent achieves superior performance on fidelity-related metrics (e.g., PSNR, SSIM). When switching to perception mode (4KAgent (P), choose to be more creative), 4KAgent achieves SOTA performance on non-reference perceptual metrics (NIQE, CLIPIQA, MUSIQ, MANIQA) with competitive performance on full-reference perceptual metrics (LPIPS, DISTS).

For AI-generated content, we report experimental results on GenAIBench-4K and DiffusionDB-4K benchmarks in Table 17 in the Appendix. We also employ SOTA generative SR methods for comparison. The experimental result based on SANA [8] is shown below:

We find that upscaling images generated by SANA-1K using 4KAgent produces better perceptual quality than directly generating 4K images with SANA-4K. 4KAgent also surpasses state-of-the-art generative SR methods in perceptual quality. As shown in Figure 10 in the main paper, SANA-1K + 4KAgent yields richer textures and better aesthetic alignment.

For old photos, Figure 5 in the Appendix shows that 4KAgent generates more realistic and detailed restorations, with finer hair strands, eyebrow patterns, and eyes compared to previous methods.

For the real-world image SR task, we conducted experiments on two commonly used real-world image super-resolution benchmarks: RealSR and DrealSR. They  are real‑world super‑resolution (SR) datasets built by capturing paired low-quality (LQ) and high-quality (HQ) images of the same static scenes using fixed DSLR cameras at different focal lengths (e.g., 28 mm for LQ vs. 105 mm for HQ). LQ images contain authentic distortions, including spatially varying blur, sensor noise, and lens artifacts, which are different from bicubic downsampling (as observed in synthetic datasets). Therefore, they are widely used by SOTA generative SR methods [4,5,6] for evaluation. Following the same experiment setting in these methods, we compare 4KAgent with SOTA generative SR methods and an agentic image restoration system. Experimental results are shown below:

Benchmark: RealSR

Benchmark: DRealSR

The experiment result leads to a similar conclusion on the traditional SR task: under the fidelity mode (choose to be more accurate), 4KAgent achieves superior performance on fidelity-related metrics (e.g., PSNR, SSIM). When switching to perception mode (choose to be more creative), 4KAgent achieves SOTA performance on non-reference perceptual metrics (NIQE, CLIPIQA, MUSIQ, MANIQA). Due to space limitations in the main paper, we provided the detailed experimental results in Table 10 in the Appendix. We will update the corresponding experimental results in the next version of the paper.

We also perform real-world remote sensing super-resolution experiments using the WorldStrat dataset (Corresponding to section F.1 in the Appendix). Each high‑resolution (HR) image in this dataset is a low-cloud Airbus SPOT 6/7 capture (1.5 m pan, 6 m RGB+NIR), while low‑resolution (LR) images consist of up to 16 Sentinel-2 revisits (10–60 m, 12–13 bands). WorldStrat thus provides a realistic benchmark for remote sensing image SR. Experiment results are shown below:

TransENet is a specialized aerial SR method trained on remote sensing data. In fidelity mode, 4KAgent performs competitively on SSIM, LPIPS, and DISTS. In perception mode, it achieves state-of-the-art results on no-reference perceptual metrics like CLIPIQA, MUSIQ, and MANIQA.

Experiments across diverse datasets show that 4KAgent is a flexible agentic system capable of producing either more faithful or more creative high-quality but arguably more creative results, depending on user preferences. We validate its superiority across traditional image SR, AI-generated image SR, and real-world natural and remote sensing image SR, with consistently strong performance.

Q2: The usage of image description in 4KAgent.

A2: In the 4KAgent system, the textual description of an image is used in the MoE module of the Restoration Agent. Specifically, when the Restoration Agent executes the restoration plan provided by the Perception Agent, it processes the input image using multiple restoration methods at each execution step, resulting in multiple candidate images. Therefore, we need to select the best image. For this purpose, we designed a quality score that considers two aspects: perception quality, which we combined non-reference IQA metrics such as NIQE, MUSIQ, MANIQA, and CLIPIQA to generate a score; and human preference, for which we used HPSv2 [9] framework to assess human preference. Since HPSv2 requires images and their corresponding descriptions for evaluation, we need to extract the descriptions of the images to evaluate human preference and assist the MoE module in making selections.

Q3: Good non reference IQA is meanful only when a comparable fidelity can be guaranteed. Provide examples of how images are corrupted and how their corresponding restore-path is.

A3: Based on our answer to Q1 & Q5 (A1 in part 1), 4KAgent can be customized to be more accurate (i.e., the traditional SR problem), or more “creative” (we think this is very useful to handle AI-generated content and old photos where more “hallucination” is needed), depending on user preferences. In both modes, 4KAgent delivers superior performance and outperforms SOTA generative SR methods on multiple image SR tasks.

For the RealSRSet experiment in our A3 in rebuttal: Since the RealSRSet dataset only contains real-world low-quality (LQ) images, we use non-reference IQA metrics for evaluation after 16x SR. Additionally, Figure 5 in the main paper provides corresponding visualization results. Compared to previous methods (SOTA classical SR methods (HAT-L), SOTA generative SR methods (DiffBIR), and the Agentic image restoration system (AgenticIR)), the images generated by 4KAgent are more faithful to the input LQ images (e.g., the rock and grass textures). Furthermore, we provide additional visualization results in Figure 5 in the Appendix, and the conclusion is the same.

We provide some specific restoration paths of 4KAgent when processing the RealSR and RealSRSet dataset (since the images in these datasets are collected from the real world, there is no process involving synthetic LQ images):

Figure 2 in the Appendix (top): Defocus Deblurring@DRBNet → Super-Resolution@PiSA-SR.

Figure 2 in the Appendix (bottom): Super-Resolution@DiffBIR → JPEG Compression Artifact Removal@FBCNN (BQF) → Defocus Deblurring@LaKDNet.

Figure 5 in the Appendix (bottom) (also the lower left image in Figure 1 with tag ‘Human Faces’): Super-Resolution@OSEDiff → Face Restoration@CodeFormer → Denoising@Restormer → JPEG Compression Artifact Removal@FBCNN (BQF) → Defocus Deblurring@DiffPlugin → Super-Resolution@DiffBIR. BQF in FBCNN indicated 'Blind to Quality Factor'.

For synthesized low-quality images and their restoration path, we provided examples on the subset of the MiO dataset:

Figure 4 in the main paper (left): containing haze and x4 downsample.

Restoration path: Super-Resolution@PiSA-SR → Dehazing@RIDCP.

Figure 4 in the main paper (right): containing rain steak and Gaussian noise, and x4 downsample.

Restoration path: Deraining@Restormer → Denoising@SwinIR(σ=15) → Super-Resolution@OSEDiff.

4KAgent successfully perceives distortions in the image and executes the corresponding restoration plan.

Q4: Consider building a real SR-agent that could analyze the blur kernel, local artifacts, and other challenges for SR.

A4: TL;DR: 4KAgent integrates nine specialized restoration methods and a customizable profile module to flexibly address diverse distortions, such as defocus blur, rain streaks, fog, noise, and more, based on user preferences and task requirements. On RealSR and DRealSR datasets, it correctly identifies and applies defocus deblurring in about 70% of cases, underscoring its accurate perception and restoration capabilities. Tested across 26 benchmarks, 4KAgent consistently outperforms both general and domain-specific super-resolution methods. Its dynamic selection of multiple SR tools per image breaks the single-model limitation, greatly expanding its real-world applicability.

4KAgent is designed for broad applicability across image restoration and super-resolution tasks, integrating methods for nine types of restoration. Additionally, following previous agentic AI systems, we equip 4KAgent with a profile module for customization based on user preferences (e.g., “super-resolve this image with high fidelity” → set ‘restore preference’ to ‘fidelity’) and requirements (e.g., “denoise the image” → set ‘restore option’ to ‘denoise’).

This enhances the usability of 4KAgent to support diverse scenarios, including not only rain, fog, but also blur and noise removal.

We validated the effectiveness of 4KAgent in real-world image SR tasks. For more details, please refer to RealSR and the DrealSR experiment in our answer to Q1 & Q5 (A1 in part 1).

Notably, the dominant distortion in the LQ images arises from blur. As both datasets collect LQ images by reducing focal length while keeping exposure, distance, and tripod fixed, the dominant blur in LQ images arises from optical defocus. Therefore, the blur in the LQ images in this dataset is closer to defocus blur. We conducted a statistical analysis of the tool chain from 4KAgent when processing images from these two datasets, specifically counting the number of times when "defocus deblurring" appeared in the tool chain. Experiment results are shown below:

4KAgent added the defocus deblurring restoration task to about 70% of the images in total, which indicates that 4KAgent accurately identifies defocus blur in LQ images and assigns the defocus deblurring task in the restoration plan. The leading perception quality in the experimental results also proves the effectiveness of this mechanism. We believe that by constructing an Agentic system, accurately identifying distortions in images, and employing advanced restoration methods, we can significantly expand the application scenarios of this system. We demonstrated the high usability of 4KAgent across 26 benchmarks across 12 tasks.

In addition to natural images, we conducted image SR experiments on AIGC images, remote sensing images, and biomedical images. Experiment results demonstrated the effectiveness of 4KAgent in SR tasks with various image domains (e.g., remote sensing images, biomedical images), including synthesis data and real-world data (e.g., WorldStrat dataset on remote sensing image SR and SR-CARO-2 dataset on fluorescence microscopic image SR). 4KAgent not only outperforms SOTA general image SR methods but also outperforms SR methods specifically designed for corresponding domains in these tasks.

Additionally, we statistically analyze the win rate of 4KAgent when processing different types of images, which refers to the top three most selected SR tools by 4KAgent when performing SR on images from this dataset, as shown in the table below:

It shows that:

(1) For different types of image SR tasks, the leading SR tools are not the same.

(2) The total number of times the leading SR tools are selected is less than the number of images in the dataset, indicating that 4KAgent uses more than three different SR tools when processing images in these datasets.

It demonstrates that 4KAgent can flexibly and appropriately invoke SR methods within the system when addressing different tasks, which breaks through the limitations of a single model and significantly enhances the applicability and application scenarios of 4KAgent.

Reference

[1] Blau, Y., & Michaeli, T. (2018). The perception-distortion tradeoff. CVPR.

[2] Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., & Timofte, R. (2021). Swinir: Image restoration using swin transformer. ICCV.

[3] Chen, X., Wang, X., Zhang, W., Kong, X., Qiao, Y., Zhou, J., & Dong, C. (2023). Hat: Hybrid attention transformer for image restoration. arXiv preprint arXiv:2309.05239.

[4] Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., ... & Dong, C. (2024). Diffbir: Toward blind image restoration with generative diffusion prior. ECCV.

[5] Wu, R., Sun, L., Ma, Z., & Zhang, L. (2024). One-step effective diffusion network for real-world image super-resolution. NeurIPS.

[6] Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., & Zhang, L. (2025). Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. CVPR.

[7] Zhu, K., Gu, J., You, Z., Qiao, Y., & Dong, C. (2024). An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809.

[8] Xie, E., Chen, J., Chen, J., Cai, H., Tang, H., Lin, Y., ... & Han, S. (2024). Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629.

[9] Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., & Li, H. (2023). Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341.

We sincerely thank Reviewer jm8i for your detailed and encouraging feedback. We appreciate your recognition of 4KAgent as a unified, domain-general framework for 4K image restoration across real photos, AI-generated content, and scientific/medical imagery. We are grateful for your acknowledgement of our two-stage Perception–Restoration Agent architecture, and the Quality-Driven Mixture-of-Experts policy with rollback. Your recognition of our extensive evaluation across various benchmarks and the consistent state-of-the-art performance across diverse and practical scenarios is especially motivating for us.

Q1: Runtime and memory measurements

A1: Thank you for pointing this out. As a multi-agent system, 4KAgent supports multi-GPU deployment. Specifically, 4KAgent assigns different agents (Perception Agent, Restoration Agent) to different GPUs to conserve memory. Most of our experiments were conducted using two NVIDIA RTX 4090 GPUs. For accelerating 4KAgent, we developed a Fast4K mode for 4KAgent. Specifically, when the image resolution is higher than the resolution threshold s set by the user (longer side > s), 4KAgent will disable long-running methods in the toolbox in order to reduce the processing time of 4KAgent. At the same time, this resolution threshold s can be set by the user, so the user can flexibly adjust it to balance performance and processing time.

Based on the Fast4K mode, we conduct an experiment on DIV4K-50 by setting different resolution threshold s (s = 1024 / 4096) in 4KAgent and evaluate the running time and performance. Experiment result is shown below:

It shows that, simply set s to 1024, the average running time of 4KAgent decreases from 1551.8s to 582.7s (62.3% ↓) with minimal effect on performance. Under this setting, 4KAgent is even faster than the model-based approach (DiffBIR). We will keep working on accelerating 4KAgent to increase its availability.

Q2: Boundary artifacts for face restoration pipeline

A2: TL;DR: The face restoration pipeline will introduce boundary artifacts in a low frequency of occurrence, and the benefits of the face restoration pipeline in 4KAgent outweigh the impact of the boundary artifacts it introduces.

Thank you for pointing this out. When designing the face restoration pipeline in 4KAgent, we follow the ‘whole image restoration’ procedure from previous methods (e.g., GFPGAN [4], CodeFormer [5], DifFace [6]). Specifically, we use the toolbox in the GFPGAN method to extract faces in the image and paste them back after restoration. Therefore, it may introduce boundary artifacts at the edge. We conduct an experiment for this.

Specifically, we use WebPhoto-Test as the testing dataset and select 100 LQ images as test images. Then, we input these images to GFPGAN, CodeFormer and DifFace for face restoration under the ‘whole image restoration’ procedure. We also collect result images from 4KAgent. After obtaining all the result images, we conduct a user study to explore the boundary artifacts. We define the boundary artifacts as: Jagged edges or color halos along the face boundary. Based on this definition, we ask users to evaluate the result images and count the occurrence of the boundary artifacts. Then we calculate the average occurrence of the boundary artifacts. Experiment results are shown below:

It shows that SOTA face restoration methods and 4KAgent introduce boundary artifacts following the ‘whole image restoration’ procedure. However, the frequency of occurrence is low (less than 10%), and 4KAgent does not increase the frequency of occurrence due to its system design.

We also conduct an ablation study of the face restoration pipeline in Section D in the Appendix. Specifically, we compare the result images from 4KAgent with different profiles (ExpSR-s4-P, ExpSRFR-s4-P) on the WebPhoto-Test dataset. The key difference between profiles and corresponding experiment results is shown below:

where ‘Face Restore’ controls the activation of the proposed face restoration pipeline. It shows that when activating the face restoration pipeline, both the overall image perceptual quality and face-related IQA metrics are enhanced. We also provide a visual comparison in Figure 8 in the Appendix, which further demonstrates the superiority of the face restoration pipeline by the enhanced facial and hair details from the image from 4KAgent with the ExpSRFR-s4-P and GenSRFR-s4-P profiles (activating the face restoration pipeline).

Therefore, following the previous procedure, the face restoration pipeline in 4KAgent will introduce boundary artifacts with a low frequency of occurrence. However, the face restoration pipeline will enhance the overall image quality and face quality. We believe that the benefits of the face restoration pipeline in 4KAgent outweigh the impact of the boundary artifacts it introduces. We will further modify the face restoration pipeline to reduce the boundary artifacts it introduces.

Q3: Expand the related work section

A3: Thank you for pointing this out. Due to space limitations, the related work section we provide in the main paper is an abbreviated version. We discuss the latest progress of SR models in detail in section B.1 of the Appendix. For example, we discussed the recent progress of generative SR methods (e.g., StableSR [7], SeeSR [8], OSEDiff [2], PiSA-SR [3]) and all-in-one image restoration methods (e.g., AirNet [9], ADMS [10]). We also discuss the recent progress of the agentic image restoration framework (e.g., RestoreAgent [11], AgenticIR [12], MAIR [13], Q-Agent [14])

At the same time, we also discuss the progress and application of SR models in Section G of the Appendix. For example, we discuss the application of SR models in video conferencing, surveillance, and gaming. We will continue to conduct in-depth research on the progress of SR models and provide a richer and more comprehensive discussion on the related work section in the next version of the paper.

Reference

[1] Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., ... & Dong, C. (2024). Diffbir: Toward blind image restoration with generative diffusion prior. ECCV.

[2] Wu, R., Sun, L., Ma, Z., & Zhang, L. (2024). One-step effective diffusion network for real-world image super-resolution. NeurIPS.

[3] Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., & Zhang, L. (2025). Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. CVPR.

[4] Wang, X., Li, Y., Zhang, H., & Shan, Y. (2021). Towards real-world blind face restoration with generative facial prior. CVPR.

[5] Zhou, S., Chan, K., Li, C., & Loy, C. C. (2022). Towards robust blind face restoration with codebook lookup transformer. NeurIPS.

[6] Yue, Z., & Loy, C. C. (2024). Difface: Blind face restoration with diffused error contraction. TPAMI.

[7] Wang, J., Yue, Z., Zhou, S., Chan, K. C., & Loy, C. C. (2024). Exploiting diffusion prior for real-world image super-resolution. IJCV, 132(12), 5929-5949.

[8] Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., & Zhang, L. (2024). Seesr: Towards semantics-aware real-world image super-resolution. CVPR.

[9] Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., & Peng, X. (2022). All-in-one image restoration for unknown corruption. CVPR.

[10] Park, D., Lee, B. H., & Chun, S. Y. (2023). All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. CVPR.

[11] Chen, H., Li, W., Gu, J., Ren, J., Chen, S., Ye, T., ... & Zhu, L. (2024). Restoreagent: Autonomous image restoration agent via multimodal large language models. NeurIPS.

[12] Zhu, K., Gu, J., You, Z., Qiao, Y., & Dong, C. (2024). An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809.

[13] Jiang, X., Li, G., Chen, B., & Zhang, J. (2025). Multi-Agent Image Restoration. arXiv preprint arXiv:2503.09403.

[14] Zhou, Y., Cao, J., Zhang, Z., Wen, F., Jiang, Y., Jia, J., ... & Zhai, G. (2025). Q-Agent: Quality-Driven Chain-of-Thought Image Restoration Agent through Robust Multimodal Large Language Model. arXiv preprint arXiv:2504.07148.

Dear Reviewer jm8i,

Thank you for your thoughtful review and encouraging rating!

As the author-reviewer discussion period is nearing its end, we would like to follow up to see if our rebuttal has addressed your concerns. Please let us know if any further clarification would be helpful.

Thank you again!

Best Regards,

Authors

We sincerely thank Reviewer APnX for recognizing the novelty of our proposed agentic framework, the comprehensiveness of our experimental validation, and the broad applicability of our method. We greatly appreciate your positive feedback on the unified design of 4KAgent, the effectiveness of our two-stage Perception-Restoration Agent architecture, the introduction of the Quality-Driven Mixture-of-Experts policy, and the dedicated face-restoration pipeline. Your acknowledgment of our state-of-the-art performance across various benchmarks, as well as the impact of our work on diverse domains such as AIGC, remote sensing, and biomedical imaging, is very encouraging.

Q1: The misalignment between line 112 / 180 and line 55

A1: Thank you for pointing this out. As we have shown in line 112 and line 180 in the main paper, 4KAgent is able to upscale an image with a scale factor of 2, 4, 8, or 16. It can achieve a larger scale factor (e.g., 32, 64, …) if applied recursively. However, it still has misalignment with ‘arbitrary-scale super-resolution’ in line 55. Therefore, we will provide a more rigorous statement in line 55 in the revised version of the paper.

Q2: Conduct twice super-resolution in perception agent and restoration agent in Figure 2

A2: In a single run, 4KAgent can upscale images with a resolution of 250 x 250 or larger to 4K resolution (the long side of the image is greater than 4000 pixels). This depends on the profile set by the user. Specifically, after perceiving the distortion of the input image and the corresponding restoration agenda, the Perception Agent will calculate the scale factor based on the resolution of the input image and the 4K resolution. The calculation method is given by the formula in line 112. The resolution of the input image in Figure 2 is 259 x 194. Therefore, the scale factor required to enlarge this image to 4K resolution is 16. To our best knowledge, SOTA super-resolution (SR) methods (e.g., SwinIR [1], HAT-L [2], DiffBIR [3], OSEDiff [4], PiSA-SR [5], …) can not directly achieve 16x SR or perform limited. We also conduct an experiment on DIV4K-50 benchmark to explore the performance of applying super-resolution (x4) twice and directly applying a 16x super-resolution. The experiment result is shown below:

It shows that applying super-resolution (x4) twice performs better than directly applying 16x super-resolution for SOTA SR methods and our 4KAgent. Therefore, we add two super-resolution (x4) tasks in the restoration agenda to achieve the goal of 16x upscaling, and then schedule the restoration plan based on the updated restoration agenda.

Q3: Correct writing errors

A3: Thank you for pointing this out. We will carefully check the writing errors in the paper and correct them in the next version.

Reference

[1] Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., & Timofte, R. (2021). Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 1833-1844).

[2] Chen, X., Wang, X., Zhang, W., Kong, X., Qiao, Y., Zhou, J., & Dong, C. (2023). Hat: Hybrid attention transformer for image restoration. arXiv preprint arXiv:2309.05239.

[3] Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., ... & Dong, C. (2024, September). Diffbir: Toward blind image restoration with generative diffusion prior. In European conference on computer vision (pp. 430-448). Cham: Springer Nature Switzerland.

[4] Wu, R., Sun, L., Ma, Z., & Zhang, L. (2024). One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37, 92529-92553.

[5] Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., & Zhang, L. (2025). Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. In Proceedings of the Computer Vision and Pattern Recognition Conference (pp. 2333-2343).