We sincerely appreciate the reviewer’s thoughtful evaluation and high-confidence endorsement of our work regarding several important aspects:

• Novelty and originality: Recognized in the summary (“novel SRSR framework”), the strengths (“novel SRCA and STCFG mechanisms offer a unique contribution,” “high quality of methodology and implementation”), and the numerical rating (“Originality: 4: excellent”).
• Comprehensive and compelling experimental results that validate effectiveness and identified research gaps: Praised for “strong empirical evidence supporting the claims,” “clear ablation studies, hyperparameter analysis, and comparisons to state-of-the-art methods,” “outperforms seven state-of-the-art baselines in fidelity and perceptual metrics on both synthetic and real-world datasets,” “sufficient and reasonable experiments,” and “Fig. 4 not only shows the superiority of the method, but also shows the limitations of existing IQAs.”
• Practical impact: Highlighted by “strong plug-and-play usability makes it practical for integration” and “the approach is fully inference-time and plug-and-play, requiring no additional training.”

Finally, we value the reviewer’s constructive feedback in the weaknesses and questions sections, and we have addressed each point thoroughly in the responses below.

W1, Q1, Q2, Q3: Generalization and robustness to segmentation and prompt extraction failures

Q3-2: “Score could increase with answers addressing generalization and robustness to segmentation or prompt extraction failures.”

We are grateful to the reviewer for explicitly outlining actionable criteria (Q3-2) that could lead to a higher rating, which is helpful and actionable for us to prepare a targeted response that can more effectively address the concern. Noting that several comments (W1, Q1, Q2, and Q3) closely relate to this point, we address them together for a cohesive and structured response.

We first clarify both intuitively and with evidence, how our proposed SRCA and STCFG are robust to errors in prompt extraction (DAPE) and segmentation masks (Grounded SAM), thereby demonstrating that our framework does not overly rely on these components (addressing W1 and Q2). We then discuss potential future directions for improving these backbones, either by replacing DAPE (Q1) or by jointly optimizing the prompt extraction and segmentation (Q3-1).

W1: “Potential over-reliance on DAPE and Grounded SAM performance, which could vary in real-world applications.”

Q2: “Could the authors clarify how sensitive SRCA and STCFG are to errors in the segmentation masks from Grounded SAM?”

(1) Robustness to failures of DAPE

For reference, the three research gaps we identify in this work (Sec. 1, lines 34-46 and Sec. 3.1, lines 109-131) are:

• G1: Diverted cross-attention
• G2: Incorrect prompts being extracted
• G3: Incomplete prompts and coverage (relevant prompts not being extracted)

G2 and G3, both stemming from DAPE’s failures, often result in hallucinations in baseline methods. Our method is specifically designed to address these, thus being robust to them:

• Robustness to G2: SRCA only retains tags that can be grounded and forms tag–mask pairs for attention re-focusing, thereby filtering out irrelevant tags wrongly extracted by DAPE. For example, in Fig. 2 (last two rows), DAPE extracts the irrelevant tag “camouflage,” causing hallucinations in the baseline, while our method removes its influence, as highlighted in the red box. Similar cases are presented in Supplementary Figs. S2 (”gun” and “rifle”) and S3 (”nose” and “mouse”).
• Robustness to G3: Without SRCA, attention from existing tags can leak into ungrounded regions (e.g., backgrounds or salient objects missed by DAPE), resulting in hallucinated content. SRCA helps suppress this, but on its own cannot prevent global prompt tokens (e.g., EOS that captures the semantics of the entire prompt, consisting of all tags) from injecting semantics into ungrounded regions. STCFG further addresses this by blocking all text-conditioning (including EOS) in ungrounded regions. An illustrative failure case is shown in Fig. S1, where DAPE fails to extract the “claw” tag (highlighted in the red bounding box), leaving the region ungrounded. In this scenario, the baseline model (without SRCA and STCFG) hallucinates this area using unrelated tags such as “animal.” In contrast, our method (with SRCA and STCFG) prevents the attention of “animals” from influencing this region, resulting in a more faithful restoration. Similar examples can be seen in Fig. S7: when DAPE misses the “plate” tag, the corresponding region remains ungrounded. Baseline approaches allow irrelevant tags to introduce artifacts in such areas, while our framework effectively mitigates these errors and preserves semantic accuracy.

(2) Robustness to failures of Grounded SAM

We also observe failure cases where Grounded SAM does not provide comprehensive masks. Our method remains robust due to the design of STCFG, which improves faithfulness in ungrounded regions. Even if coverage is partial, SRCA ensures higher fidelity in grounded areas compared to the baseline. A visual example is shown in Fig. S1, where Grounded SAM fails to cover both “stone” objects in the image fully, and only the lower stone is confidently grounded, leaving a substantial ungrounded region. In the grounded area (yellow box), our method achieves higher fidelity. While the baseline hallucinates small pebble-like textures, our use of SRCA ensures that only the tag “animal” (and not “stone”) influences restoration, yielding more accurate results. For the ungrounded stone regions, STCFG effectively prevents over-synthesized or artifact-prone textures that the baseline exhibits, resulting in reconstructions much closer to the ground-truth HR image. Similar examples can be seen in Fig. 1 (Left). When the bird’s body is not fully grounded, STCFG prevents the introduction of artifacts from irrelevant tags such as “stone,” preserving the integrity of the ungrounded region.

Additionally, our comprehensive hyperparameter analysis in Sec. 4.4 and Tab. 3 demonstrate that our method is robust to varying thresholds for Grounded SAM, consistently outperforming the baseline.

(3) Potential directions

Q1: How does the performance change if the prompt extractor (DAPE) is replaced with a newer vision-language model or fine-tuned for specific domains?

Q3-1: Have the authors considered joint optimization between segmentation and SR to resolve interdependencies more robustly?

We appreciate these insightful suggestions. Following the baselines of SeeSR and OSEDiff, DAPE is adopted due to its proven advantage in Real-ISR in terms of degradation-awareness and efficiency, outperforming other off-the-shelf models like ResNet, YOLO, BLIP, and LLaVA. This is because DAPE is already fine-tuned on the Recognize Anything Model (RAM) for this domain, as suggested in Q1.

Nevertheless, as identified in G2 and G3 and previously discussed, there remains room for improvement. As discussed in the future work section (Supplementary Sec. E), we believe that advances in the prompt extraction and segmentation backbones can further improve SRSR’s performance. This is grounded on our observation from ablation studies that stronger semantic segmentation backbones (e.g., DINO-X over Mask2Former) lead to better results. Also, in line with Q3-1, we believe joint optimization of segmentation and SR is a promising and novel direction that could further enhance the current backbones. We are grateful that the reviewer endorses our comprehensive experiments and compelling results, which show that our current DAPE + Grounded SAM backbones already achieve consistent state-of-the-art results. As new advances emerge in these backbones, we believe our method will benefit from orthogonal improvements.

W2: User study

No user study or real-world deployment evaluation to validate perceptual benefits from an end-user perspective. Fig. 4 displays the limitations of existing IQAs, which further indicates the necessity of user study to evaluate the motivation.

Thank you for highlighting the importance of evaluating perceptual benefits from an end-user perspective. In response to this motivation, we have conducted a user study to evaluate human perceptual preferences between the baseline results and those produced after integrating SRSR, details of which are included in Supplementary Material Sec. D and Fig. S18. In our study, we presented the low-resolution input for each image along with two anonymized high-resolution outputs (ours: SRSR-SeeSR and the baseline: SeeSR) labeled ‘Method A’ and ‘Method B’. Annotators were asked to select the preferred image or indicate if both appeared visually similar, with evaluations based on sharpness, visual realism, and detail preservation. The order of presentation was randomized, and full images were shown (not crops) to avoid bias. Results are aggregated and visualized in bar charts (Figure S18), showing that SRSR-SeeSR is consistently preferred over the baseline SeeSR across both datasets.

We sincerely thank the reviewer for their detailed evaluation and positive assessment of our work. We are especially encouraged by the reviewer’s recognition of our paper’s strengths across all four criteria: quality, clarity, originality, and significance:

• Motivation and identified research gaps: “The paper is well-motivated, clearly identifying critical and unsolved issues in SD-based Real-ISR frameworks.”
• Novelty and methodological elegance: “The proposed solutions are elegant,” and “STCFG introduces a novel use of classifier-free guidance in Real-ISR.”
• Comprehensive and convincing experiments: “The experiments are comprehensive, and the qualitative results are particularly convincing,” “Experiments demonstrate consistent improvements in image fidelity,” and “The illustrative examples and experiments tightly support the claims.”
• Efficiency and practical merits: “The plug-in framework enhances Real-ISR models without the need of retraining, making it efficient and practical,” and “SRSR is compatible with existing models without retraining.”
• Clarity and technical soundness: Recognized by the numerical rating (“Clarity: 4: excellent”) and the encouraging comment: “This paper is technically sound and clearly written.”

Finally, we appreciate the reviewer’s constructive feedback in the weaknesses and questions sections, and we address each point thoroughly in the responses below.

W1, Q2, Q3, and Limitation: Source of improvement and limitation of no-reference metrics

Given the close connections among W1, Q2, Q3, and the Limitation section, we address them together to provide a cohesive and structured response. We first disentangle the source of improvement between SRCA and STCFG (addressing W1 and Q3), and then offer further insights regarding the limitations of current no-reference metrics and the broader perception–fidelity trade-off (addressing Q2 and Limitation, which are also relevant to W1 and Q3).

(1) Source of improvement: SRCA vs STCFG

W1: Improvements may largely stem from simply reducing classifier-free guidance (CFG) through the proposed STCFG, rather than targeted semantic enhancement.

Q3: Source of improvement: SRCA vs STCFG.

We appreciate the reviewer’s thoughtful comments and agree that STCFG is the main driver of performance gains. However, we have noticed a slight but crucial misconception in W1 comments; thus, it is important to clarify that the improvements from STCFG are not independent of targeted semantic enhancement—rather, both SRCA and STCFG contribute synergistically to this goal (instead of SRCA solely contributing towards this). The reason why STCFG is more significant, in a nutshell, is that it effectively addresses the critical and widely observed shortcomings of grounding and SRCA (w/o STCFG) that are otherwise detrimental to semantic accuracy. Thus, it is important to highlight that such improvements are also for semantic accuracy. We will elaborate in the following paragraphs.

SRCA refines semantic conditioning via more precise cross-attention maps, but in ungrounded areas, its effect in ungrounded areas is limited by the persistence of summary tokens (e.g., EOS) that capture semantics of all tags in the prompt (including the mismatched tags) and meaningless separator tokens, which can inadvertently introduce mismatched or noisy semantics. Note that masking these tokens is not feasible, as it disrupts the diffusion model’s noise prediction.

Moreover, grounding (required for SRCA) reduces prompt coverage and increases ungrounded regions, especially for degraded LR images; thus, although SRCA alone has clear benefits, substantial room for improvement remains.

STCFG addresses the limitations of SRCA by replacing text-conditioned noise with unconditional noise in ungrounded regions, reducing hallucinations while preserving SRCA’s benefits.

Lastly, regarding the comparisons with segmentation-based methods in our ablation studies (V7 and V8 in Tab. 2), we wish to clarify that both versions already incorporate STCFG to ensure fair comparison with our final choice. Removing STCFG (see additional results: V9, V10) further degrades performance, highlighting its importance. With semantic segmentation (SS), the increased coverage reduces the frequency of ungrounded regions, so STCFG is less often triggered and brings smaller additional benefit than the setting without SS.

(2) Limitation of no-reference metrics and the Perceptual-Fidelity trade-off

Q2: Perception–fidelity trade-off.

Limitation: No-reference metrics.

We appreciate the reviewer’s acknowledgment that this remains a challenging and open question in Real-ISR. The drop in no-reference metric scores is due to the same improvements in semantic accuracy (i.e., reduced hallucinations). No-reference metrics tend to over-reward artifact-heavy or hallucinated outputs, as noted in our paper and acknowledged by the reviewer. In contrast, full-reference metrics we employed also have dedicated measures for evaluating perceptual quality (LPIPS and DISTS) and fidelity (PSNR and SSIM). Our improvements in these metrics are well-aligned with the visual evidence we provide in the paper, such as Fig. 4, and Supplementary Material Figures S1-S17.

In SUPIR (Fig. 7), human preference can favor creative outputs diverging from the reference, but in Real-ISR, especially for applications needing semantic faithfulness (e.g., e-commerce product images), hallucinations reduce trust and utility. Thus, while SD-based SR models excel in producing creative results that are visually pleasing, an important question is that shall we reward creativity in Real-ISR as much as possible by pursuing favorable no-reference metrics results? Our view is that faithful restoration, as measured by full-reference fidelity and perceptual quality metrics, should remain a central goal in many practical contexts. For V2–V4, we focus on full-reference metrics: V3 (SRCA only) improves fidelity but reduces perceptual quality, while V4 (SRCA+STCFG) significantly improves both, aligning with qualitative evidence.

Path forward: We fully agree that in scenarios lacking references, reliance on existing no-reference metrics is problematic, and developing improved, hallucination-aware no-reference metrics that better align with full-reference scores and human perception is an urgent research need. Promising directions include vision-language model-based assessment and metrics penalizing hallucinations. Nonetheless, in the standard settings of Real-ISR and similar settings when references are available, we advocate for full-reference metrics as a more complete and reliable evaluation. We appreciate the reviewer’s feedback and will incorporate these points in our revised Limitations section.

W2: The contribution of grounding

Thanks for raising this point. While grounding improves prompt accuracy by removing redundant tags, it also reduces coverage, increasing ungrounded regions. Thus, the modest gain from V1 to V2 reflects a suboptimal trade-off. Our STCFG module compensates for this limitation by improving quality in ungrounded regions. As a result, SRSR is robust to grounding thresholds (Tab. 3), consistently significantly outperforming the baseline.

We also wish to emphasize that grounding's contribution extends significantly beyond the performance improvements from V1 to V2. Crucially, it provides the essential tag–mask pairs necessary for enabling our SRCA and STCFG modules, without which these key innovations would not be feasible.

Q1: Motivation clarity

Thank you for this question. Our conclusion that “DAPE is only partially robust to severe degradations” is based on failure case analysis, empirical observations, cross-attention map analysis, and quantitative comparisons.

For reference, our study identifies three main research gaps (Sec. 1, lines 34–46 and Sec. 3.1, lines 109–131):

• G1: Diverted cross-attention
• G2: Incorrect prompts being extracted
• G3: Incomplete prompts and coverage (relevant prompts not being extracted)

We observed hallucinated objects in the baseline linked to DAPE-extracted tags not grounded in the image (G2). Cross-attention maps confirmed their impact. We also found that DAPE sometimes misses relevant tags, leading to ungrounded regions and further hallucinations (G3). This is supported by qualitative and quantitative evidence throughout the paper and the supplementary materials. For example:

• G2 (Incorrect tags): In Fig. 2 (last two rows), DAPE extracts the irrelevant tag “camouflage,” leading to hallucinations in the baseline. Similar cases appear in Supplementary Figs. S2 (“gun” and “rifle”) and S3 (“nose” and “mouth”), etc.
• G3 (Incomplete tags): In Fig. S1, DAPE fails to extract the “claw” tag, resulting in an ungrounded region where the baseline hallucinates this area based on unrelated tags such as “animal.” A similar pattern occurs in Fig. S7, where missing the “plate” tag allows irrelevant tags to introduce artifacts.

Q4: Incorporating SRSR into training

We appreciate the reviewer’s interest in training-time integration of SRSR, which we briefly mentioned in our Future Works section. Our intuition is that incorporating grounding, SRCA, and STCFG during training would align the model’s learning with the spatial and semantic constraints imposed at inference, improving region-specific supervision and semantic localization. Integrating STCFG in training would also improve robustness to incomplete or noisy tags. We consider this a promising direction for future work.

Thank you very much for your careful review of our rebuttal and for these detailed follow-up questions. We are grateful to hear that our rebuttal has successfully addressed your concerns, and our additional insights on the limitations of existing no-reference metrics and their path forward are helpful. We are happy to address your insightful follow-up questions comprehensively one by one, as follows:

Follow-up Q1: Clarification of “STCFG addresses the limitations of SRCA by replacing text-conditioned noise with unconditional noise in ungrounded regions.”

Thank you for raising this important point. We would like to clarify that the quoted explanation in our rebuttal is technically equivalent to “STCFG disables classifier-free guidance in ungrounded regions” (as stated in Lines 51–54 of the paper). The core idea is captured in our noise prediction equations, as detailed below.

Formally, as shown in Eq. 6 of the paper, the noise prediction at each pixel $i$ is defined as:

\hat{\epsilon}_i \leftarrow (1 - M_i)({\epsilon}_θ(x_t, \phi) + s ({\epsilon}_θ(x_t, y)-{\epsilon}_θ(x_t, \phi))) + M_i {\epsilon}_θ(x_t, \phi) \tag{6}

where $M$ is the ungrounded mask for each image, and $M_i \in \{0,1\}$ indicates whether pixel $i$ is ungrounded.

• For grounded regions ($M_i=0$), the equation simplifies to the standard classifier-free guidance (CFG) formulation, where the noise prediction leverages both unconditional and text-conditional branches.

\hat{\epsilon}_i \leftarrow {\epsilon}_θ (x_t, \phi) + s ({\epsilon}_θ(x_t, y) - {\epsilon}_θ (x_t, \phi)) \tag{6a}

• For ungrounded regions ($M_i=1$), the equation becomes only using the unconditional noise prediction, while the text-conditional term is fully removed:

\hat{\epsilon}_i \leftarrow {\epsilon}_θ (x_t, \phi) \tag{6b}

Comparing these two forms, it is clear that for ungrounded regions (Eq. 6b):

• STCFG replaces the text-conditioned noise $\epsilon_\theta(x_t, y)$ with unconditional noise $\epsilon_\theta(x_t, \phi)$ in Eq. 6a, resulting in the second term being eliminated rather than just decreasing the guidance strength, becoming Eq. 6b. This represents the explanation in our rebuttal that “STCFG addresses the limitations of SRCA by replacing text-conditioned noise with unconditional noise in ungrounded regions.”
• This directly corresponds to “disabling classifier-free guidance” of Eq. 6a (as we mentioned in Lines 51-54), as the guidance (the difference between conditional and unconditional branches) is set to zero. As a result, Eq. 6b represents how diffusion models conduct unconditional generation (i.e., disabling the text-conditional generation capability provided by the second term of CFG).

We hope this clarifies that both descriptions are fully consistent. We will ensure this equivalence is clearly stated in our revision. Thank you again for giving us the opportunity to clarify this technical point.

We sincerely thank the reviewer for recognizing several important strengths of our work in the Summary section, including:

• the plug-and-play design of our spatially re-focused cross-attention module,
• our novel discovery that traditional (vanilla) cross-attention induces hallucinations by diverting attention toward irrelevant pixels, and
• the effectiveness of our strategy in mitigating interference from unrelated pixels.

We further appreciate the reviewer highlighting in the Strengths section that, based on our aforementioned novel discovery/reveal of a critical research gap, our proposed method can:

• visibly alleviate hallucinations, and
• significantly improve the fidelity of the super-resolved images.

We also appreciate the valuable feedback listed in the Weaknesses section, and we address them thoroughly in the responses below.

W1: Novelty

We appreciate the reviewer for raising this critical point about novelty. We are encouraged that the reviewer’s originality rating is “3: good,” and that all other reviewers gave either “3: good” or “4: excellent,” with explicit remarks such as “novel SRCA and STCFG mechanisms offer a unique contribution,” “STCFG introduces a novel use of classifier-free guidance in Real-ISR,” and “novel SRSR framework.”

We respectfully wish to clarify a key misunderstanding: our framework goes substantially beyond simply using segmentation masks as additional priors. Before elaborating on our specific novel contributions, we wish to provide a high-level summary that the novelty of our approach lies in how we (1) design and (2) deploy the spatial prior, addressing several under-explored and fundamental research gaps in text-conditioned Real-ISR.

• Design motivation: Rather than using off-the-shelf segmentation masks as generic priors, our masks are constructed through grounded degradation-aware tags, and our entire framework is motivated by three systematically identified research gaps (Sec. 1, 3.1):
  • G1: Diverted cross-attention (Novel finding that traditional cross-attention leads to hallucinations by diverting attention to irrelevant pixels.)
  • G2: Incorrect prompts being extracted (Although many works attempt to develop better text extraction methods for Real-ISR, we identify that even state-of-the-art, degradation-aware prompt extractors can still frequently output incorrect tags due to heavily degraded LR inputs.)
  • G3: Incomplete prompt coverage.
• Deployment: Rather than simply adding segmentation as a conditioning signal, our modules refine the core cross-attention and CFG mechanisms, changing where and how the model focuses semantic information. This approach fundamentally differs from previous work and is specifically aimed at enhancing semantic accuracy, a largely under-explored challenge.

Specific Novel Contributions

• SRCA (Spatially Re-focused Cross-Attention) directly addresses G1 and G2:
  • G1: To our knowledge, no prior work has explicitly re-designed the cross-attention mechanism itself to resolve its inherent limitations. Prior methods typically focus on improved text prompts or additional conditioning, while continuing to use default cross-attention. Our work is motivated by our novel observation that unrefined cross-attention causes semantic hallucinations—an insight backed by new empirical evidence (see ablation and visualizations).
  • G2: Instead of similarly exploring better prompt extractors, we are the first to use the idea of grounding in this task to improve prompt accuracy further.
• STCFG (Spatially Targeted Classifier-Free Guidance) addresses G3:
  • Simply using segmentation masks to guide restoration (as the reviewer assumed to be our design) can improve coverage; we have actually conducted ablation studies (Tab. 2) to prove the inferiority of this design. We have gone several extra miles, such as analyzing the limitations of using SRCA alone and other ablated versions. These motivated us to design STCFG, a novel refinement of CFG tailored to Real-ISR. It dynamically disables CFG in ungrounded regions, solving the accuracy-completeness trade-off, and leading to joint improvements in fidelity and perceptual quality.

Summary

In summary, our work is not about adding segmentation as a prior, but about re-thinking how spatial priors can be designed and integrated to solve specific, under-addressed issues in text-conditioned super-resolution. These deliberate design choices (guided by empirical discovery of new gaps) underpin the novelty and effectiveness of our SRSR framework. While our modules are deliberately simple and efficient, we believe their targeted, insight-driven design is what makes them both practical and genuinely novel.

W2: Length of Writing

We appreciate the reviewer’s suggestion to further improve the writing by ensuring that each section is appropriately proportioned. We are committed to making the paper as accessible and polished as possible. In response to this helpful feedback, we will revise the abstract to provide a more complete overview and streamline the first paragraph of the introduction to improve flow and readability.

“Existing Stable-Diffusion-based super-resolution approaches often exhibit semantic ambiguities due to inaccuracies and incompleteness in their text prompts, coupled with the inherent tendency for cross-attention to divert towards irrelevant pixels. These limitations can lead to semantic misalignment and hallucinated details in the generated high-resolution outputs. To address these, we propose a novel, plug-and-play spatially re-focused super-resolution (SRSR) framework that consists of two core components: first, we introduce Spatially Re-focused Cross-Attention (SRCA), which refines text conditioning at inference time by applying visually-grounded segmentation masks to guide cross-attention. Second, we introduce a Spatially Targeted Classifier-Free Guidance (STCFG) mechanism that selectively bypasses text influences on ungrounded pixels to prevent hallucinations. Extensive experiments on both synthetic and real-world datasets demonstrate that SRSR consistently outperforms seven state-of-the-art baselines in standard fidelity metrics (PSNR and SSIM) across all datasets, and in perceptual quality measures (LPIPS and DISTS) on two real-world benchmarks, underscoring its effectiveness in achieving both high semantic fidelity and perceptual quality in super-resolution.”

We believe this revised abstract better balances clarity and completeness, and remain open to further improvements based on feedback.

W3: Explanation of Figure 1

We thank the reviewer for this observation. The regions not highlighted in the attention matrix correspond to areas that could not be confidently grounded by Grounded SAM using the input tag (in this case, “bird”). Since Grounded SAM operates directly on the low-resolution (LR) input, degradation in the image can cause it to be more conservative in assigning segmentation masks, resulting in certain relevant regions (such as the bird’s body) being left ungrounded.

We refer to such areas as ungrounded regions throughout the paper. We wish to highlight and reassure the reviewer that we have already analyzed this limitation in Section 3.3 and specifically designed STCFG to bypass text influence in ungrounded regions, preventing hallucinations and reducing the negative impact of incomplete coverage.

W4: Additional References

We thank the reviewer for noting these related works, and will discuss both in the revision:

• SFT demonstrates how spatially-varying categorical priors can be integrated into SR networks to recover more realistic textures.
• CoSeR highlights the potential of bridging image and language understanding using diffusion-based priors and cognitive embeddings.

While these methods use semantic or language guidance, our SRSR framework introduces a grounded and spatially-refocused mechanism (SRCA + STCFG) that directly tackles the challenge of semantic ambiguities caused by diverted cross-attention, inaccurate text prompts, and incomplete/ungrounded text prompts, which are not explicitly addressed by SFT or CoSeR.

W5: Why not use open-vocabulary panoramic segmentation models?

We appreciate this suggestion. In fact, we independently explored this exact idea and incorporated it into our work. Specifically, in Sec. 3.3, Sec. 4.3, and Supplementary Sec. B (Figures S12 and S13), we quantitatively and qualitatively evaluated Prompt-Free Anything Detection and Segmentation from DINO-X, a state-of-the-art open-vocabulary segmentation method.

While this approach improves coverage over Mask2Former, its lack of degradation-awareness compromises accuracy on degraded LR inputs, leading to incorrect prompt-region associations and inferior SR performance (see Tab. 2, V8 vs. V4). Supplementary Figs. S12–S13 further illustrate these issues, which our method overcomes. We will clarify this in the revision.

W6-W8: Minor writing and formatting issues

Thank you for these helpful observations. We will address all the mentioned concerns in the revision:

• W6: We will revise the methodology sub-sections accordingly by highlighting the strength of our proposed SRCA and STCFG modules before the detailed explanations.
• W7: We will revise the tense in line 179 to maintain consistency.
• W8: We acknowledge the duplicated references (9 and 10, 45 and 46) and will remove the redundancies in the updated bibliography.

We sincerely thank the reviewer for their thoughtful evaluation and positive assessment of our work regarding several important aspects:

• Comprehensive experiments and effective results: “SRSR improves both visual quality and alignment with semantics and also alleviates hallucination, outperforming baselines across multiple datasets and full-reference evaluation metrics,”
• Practical merits: “The experiment results demonstrate that the proposed method can be easily plugged into existing text-conditioned Stable-Diffusion-based SR methods and improve the SR quality with minimal additional burden,” and “a plug-and-play and training-free framework that can be easily applied to Stable-Diffusion-based SR methods with text conditions.”
• Intuitive and original methodology: Recognized by the numerical rating (“Originality: 3: good”) and the encouraging comments: “The proposed method is simple, straightforward, and intuitive,” and “the proposed method enhances semantic fidelity by grounding tags to guide cross-attention and suppress irrelevant text influences on ungrounded regions.”
• Clarity and presentation: “Overall, the writing and figures are clear and enjoyable to read.”

Finally, we appreciate the reviewer’s constructive feedback in the weaknesses and questions sections, and we have addressed each point thoroughly in the responses below.

W1: Why is our framework degradation-aware?

Thanks for this important concern regarding degradation-awareness. We would like to clarify below the fundamental difference between our proposed framework and the alternative semantic segmentation-based designs (Mask2Former and Prompt-Free Anything Detection and Segmentation in DINO-X), highlighting the key reasons for our superior results.

While we acknowledge that Grounded SAM itself is not inherently degradation-aware, our framework introduces degradation-awareness explicitly at the prompt extraction stage (prior to employing Grounded SAM). Specifically, we utilize a Degradation-Aware Prompt Extractor (DAPE), which is carefully fine-tuned from the Recognize Anything Model (RAM) to ensure extracted text tags are robust and accurate even under degraded, low-resolution (LR) conditions. Consequently, when these degradation-aware tags from DAPE are provided as inputs to Grounded SAM, the resulting prompt–mask pairs inherently exhibit enhanced degradation-awareness and accuracy. In contrast, semantic segmentation methods directly assign segmentation labels as prompts without a separate degradation-aware prompt extraction step, thus limiting their robustness and accuracy for degraded LR images.

To substantiate this claim, we also conducted comprehensive analyses:

• Quantitative ablation studies (Sec. 4.3, Table 2): Variants V7 (Mask2Former) and V8 (DINO-X), both lacking degradation-awareness, show clearly inferior results compared to our proposed approach (V4).
• Qualitative analysis (Supplementary Sec. B, Figures S12–S15): For example, Figures S12 and S13 demonstrate how the lack of degradation-awareness in semantic segmentation methods causes incorrect yet confidently-extracted prompts, subsequently misguiding restoration and producing semantically incorrect outputs. In contrast, our proposed DAPE + Grounded SAM framework effectively addresses this limitation, yielding substantially improved semantic correctness.

We will ensure these clarifications are further emphasized in the revised version.

W2: Grounding and Accuracy-Completeness Trade-off

We appreciate the reviewer’s thoughtful comment regarding grounding and the accuracy-completeness trade-off. Below, we unpack this issue into three parts to comprehensively address your concern:

1. The importance of the accuracy-completeness trade-off;
2. Why accuracy typically outweighs completeness;
3. The broader contribution of grounding beyond direct performance improvement.

(1) Importance of accuracy-completeness trade-off:

Accuracy and completeness of the extracted prompts are two of the three key research gaps we explicitly identified in Section 3.1. In our preliminary exploration (Table 2, comparing V1 to V2), we found grounding to be a promising direction as it can provide slight improvements to the baseline by simply applying grounding to improve on accuracy at the cost of completeness (without integrating deliberately designed modules such as SRCA and STCFG). At this stage, we also observed that grounding alone yields only small improvements, as the accuracy gains come at the cost of completeness, resulting in modest net gains. This clearly highlighted the necessity of explicitly addressing the accuracy-completeness trade-off. This motivated our design of the STCFG module, which significantly improved this trade-off and overall restoration.

(2) Why accuracy typically outweighs completeness:

The importance of accuracy over completeness is much more significant after employing STCFG to address the aforementioned important challenge of the accuracy-completeness trade-off. Our core insight is that reduced completeness mainly affects ungrounded regions. STCFG targets these specifically, reducing issues from incomplete coverage and making the completeness a much less concerning issue than incorrectness.

After integrating STCFG (V6), we observed substantial gains compared to V1, showing that optimized accuracy outweighs losses in completeness. Further evidence supporting this claim comes from comparing our DAPE + Grounded SAM approach (V4) against segmentation-based methods (V7 and V8). Although segmentation methods inherently achieve higher completeness, their less accurate (due to poorer degradation-awareness) prompts directly led to inferior overall performance relative to our method.

(3) Contribution of grounding beyond direct performance improvement:

Finally, we emphasize that grounding's contribution extends significantly beyond the performance improvements from V1 to V2. Crucially, grounding provides the essential tag–mask pairs necessary for enabling our SRCA and STCFG modules, without which these key innovations would not be feasible.

Q1: Implementational detail

We appreciate the opportunity to clarify the implementation of the ablated version V5. V5 first retains the entire settings of V4 and builds on it by additionally including ungrounded tags to demonstrate their negative impact. Specifically, V5 retains the SRCA refinement for grounded regions, where tag–mask pairs are available, and continues to use STCFG to block text influence in ungrounded regions. The only change is that, for V5, in addition to V4, we also keep the ungrounded tags and allow their default (inherent) cross-attention maps, without applying SRCA, since these tags do not have associated segmentation maps (as the reviewer correctly noted).

Because STCFG continues to restrict text influence in the ungrounded regions, the effect of these ungrounded tags is limited to their flawed cross-attention behavior within the union of all grounded regions. This setup highlights the drawback of including ungrounded tags: without the benefit of SRCA, their attention may be diffused or misdirected across the object regions, as seen for the “camouflage” tag in the bottom example of Fig. 2 that caused the baseline’s hallucinated object (highlighted in red), where the SRCA map would otherwise be entirely black and help effectively avoid the hallucination.

Q2: Application to other works

Thank you for this insightful question. We have not yet applied SRSR to AddSR or TSD-SR, but due to the plug-and-play nature of our framework, SRSR can be flexibly integrated into these and other recent approaches to further enhance their performance. In this work, we demonstrated the generality and effectiveness of SRSR by incorporating it into both SeeSR and OSEDiff, and we observed consistent improvements across multiple benchmarks and evaluation metrics. We will include a discussion in the revised manuscript (Future Work section) to highlight the potential for SRSR to benefit existing methods like AddSR, TSD-SR, and future methods. We believe the community will benefit from broader adoption of SRSR, and we encourage future research to explore its integration with a wider range of Real-ISR frameworks.

Q3: Different diffusion steps

Thank you for this valuable question. We would like to clarify that the greater benefit observed for SeeSR, compared to OSEDiff, primarily results from the use of STCFG rather than differences in the number of diffusion steps. Specifically, in OSEDiff, CFG is disabled by default, which renders STCFG inapplicable, meaning the performance gain comes solely from SRCA. In contrast, SeeSR benefits from both SRCA and STCFG, leading to more substantial improvements.

Motivated by the reviewer’s insight, we conducted additional controlled experiments to further evaluate the effect of diffusion steps on performance, keeping all other factors fixed. For both our method and the baseline, we compared results at 50, 40, and 30 sampling steps. As shown in the table below, while reducing the number of diffusion steps introduces the typical fidelity–perceptual quality trade-off (i.e., fewer steps improve on fidelity metrics PSNR/SSIM but worsen perceptual quality metrics LPIPS/DISTS for both methods), our method consistently maintains a significant and stable advantage over the baseline across all metrics and step choices.