acceptodds
Under review as a conference paper at ICLR 2027

Sketch Before Pixel: Pixel-Space Structure Forcing for Real-World Super-Resolution

Abstract

Pixel-space diffusion model offers a promising alternative for Real-world Image Super-Resolution (Real-SR), enabling end-to-end restoration without lossy VAE compression. However, existing pixel-space generation methods typically couple structural reconstruction and detail generation along a single denoising trajectory, making joint optimization more challenging for Real-SR. To address these limitations, we propose Sketch2Pix, a VAE-free Pixel-Space Diffusion Transformer that restores images from structural sketches to fine-grained details. Specifically, we first discover that performing moderate structural restoration before image refinement can improve the restoration quality. Based on this discovery, we design a Dual-Time Joint Diffusion framework that leverages the low-dimensional manifold hypothesis of natural images to jointly predict clean low-frequency structure and image, thereby enabling end-to-end pixel-space restoration. Furthermore, to improve the quality of the recovery, we devise a Structure-Forcing Schedule strategy that reorders the structure and pixel denoising trajectories to explicitly control the recovery order, allowing structure to emerge earlier and guide subsequent pixel generation. Benefiting from the above designs, our Sketch2Pix preserves complete pixel information while transforming coupled restoration into an ordered structure-to-detail process. Extensive experiments on Real-SR benchmarks illustrate that our Sketch2Pix restores perceptually realistic details while faithfully preserving structural content and enabling efficient inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Real-World Super-Resolution, Pixel-Space Diffusion

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SeubOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Sketch2Pixel, a VAE-free image restoration diffusion model that recovers a low-frequency structural representation earlier than the full image, allowing progressively restored structure to guide detail generation. In detail, it proposes Dual-Time Joint Diffusion, which jointly models the image and a wavelet-derived low-frequency structure with separate diffusion times, a shared DiT backbone, and two prediction branches. Besides, it introduces a structure-forcing schedule to control relative denoising progress. Experiments show advancement of the proposed model in performance and efficiency trade-offs.

Critical strengths
  • The key concept of "Sketch Before Pixel" is intuitive and telling, and the overall performance surpasses existing methods to some extent.
Critical weaknesses
  • The innovation is incremental and most concepts came from latent forcing. The previous work has already validated a similar schema that models pixels and auxiliary representations (including downsampled pixels) in the generation task. More precisely, it seems an extension of latent forcing in the restoration task, providing little insight.
  • The motivation is intuitive but vague, lacking broader pre-validation. The structure-pre schema makes sense, but it just selects the wavelet operation without explaining "why wavelet should be selected", nor providing proof that it outperforms other structural or semantical representations.
  • Despite mentioning previous VAE-free work in the image restoration task, e.g., PixelIR and PixRestore, it provides none detailed comparison or discussion.
Questions for authors

See Weaknesses.

9 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

We would like to sincerely thank the reviewer for the detailed and constructive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.

Weakness 1: Distinctions from Latent Forcing and Insights into Real-World Super-Resolution

Answer: Thank you for your insightful comment. We would like to clarify our contributions in terms of Motivation, Method, and New Insight for Real-SR.

(1) Motivation. Latent Forcing [1] studies the coordination of auxiliary representations and pixel generation. In contrast, we investigate how to restore and exploit the imperfect structural cues already present in degraded observations to constrain detail synthesis. We therefore treat structure as an explicit recovery objective that continuously updates during joint denoising, providing progressive guidance for detail reconstruction.

(2) Method. Compared with the representation configurations in Latent Forcing that rely on pretrained encoders, we adopt a computationally efficient, learning-free wavelet transform to construct spatially aligned low-frequency structural targets. Conditioned on the LR input, we jointly restore structure and the full image while independently controlling their denoising progress, forming a self-contained pixel-space restoration framework without an external encoder.

(3) New Insight for Real-SR. We find that jointly restoring structure and coordinating its progress with detail synthesis improves Real-SR reconstruction compared with conditioning on fixed structural cues from degraded LR images. Specifically, Figure 6 demonstrates the benefits of joint structural recovery over fixed LR structural conditioning. Figure 4 and Table 4 further show that moderate structural precedence improves reconstruction, whereas excessive precedence degrades quality. Motivated by these findings, we propose the Sketch Before Pixel paradigm, which explicitly coordinates structural recovery and detail synthesis within joint denoising to balance structural fidelity and visual realism.

Weakness 2: Motivation for Wavelet-Based Structural Representation

Answer: Thank you for your constructive comment. We would like to clarify that we choose wavelets to construct a spatially aligned structural target that captures coarse-scale layout and can be explicitly restored. Wavelet decomposition explicitly separates low-frequency structure from high-frequency details, allowing the structural branch to guide the coarse spatial layout and the full-image branch to recover fine details, consistent with our design of coordinating structural recovery and detail synthesis through their relative denoising progress. Moreover, this representation requires neither learning nor an external pretrained encoder and incurs little computational overhead.

Following your suggestion, we have added an ablation study replacing the wavelet-based structural target with Gaussian low-pass filtering. As shown in Table R1, the LL-based wavelet representation achieves better PSNR, SSIM, LPIPS, MUSIQ, MANIQA, and DeQA on ScreenSR, whereas Gaussian low-pass filtering performs better in NIQE and FID. These results provide empirical support for our choice of a wavelet-based structural representation. The corresponding results and discussion have been added to the revised manuscript.

Table R1. Ablation of structural representations on ScreenSR. Bold indicates the better result.

Structural representationPSNR↑SSIM↑LPIPS↓NIQE↓MUSIQ↑MANIQA↑DeQA↑FID↓
Gaussian low-pass filtering23.320.65520.24203.7764.740.32013.0757.99
Wavelet (LL, Ours)23.680.68170.20963.8071.110.50683.8162.53

References:

[1] Alan Baade et al. Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation. International Conference on Machine Learning (ICML), 2026.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 3: Comparison with VAE-Free Restoration Methods

Answer: Thanks for your valuable comment. As suggested, we have added quantitative and qualitative comparisons with PixRestore [2] under a consistent evaluation protocol. Since PixRestore was trained on part of ScreenSR and its training/test split differs from ours, we exclude direct comparisons on this dataset to avoid potential overlap between its training data and our test data.

(1) Quantitative Comparison. As shown in Table R2, PixRestore achieves higher PSNR and SSIM, but this fidelity advantage comes with poorer perceptual quality, as reflected by all four no-reference metrics. Specifically, on RealSR and LSDIR, our method improves MUSIQ by 12.06 and 7.51 points and MANIQA by 0.1991 and 0.2170, respectively. These results support the perceptual quality advantages of our method, while highlighting the trade-off between visual realism and structural fidelity.

(2) Qualitative Comparison. Compared with PixRestore, our method more faithfully preserves fine textures and text structures while recovering more realistic details, as illustrated by the examples in Figure 5 of the revised manuscript.

We did not include a comparison with PixelIR because its source code was not publicly available. The complete comparison results and accompanying discussion have been added to the revised manuscript.

Table R2. Quantitative comparison with PixRestore on RealSR and LSDIR. Bold indicates the better result.

DatasetMethodPSNR↑SSIM↑LPIPS↓NIQE↓MUSIQ↑MANIQA↑DeQA↑FID↓
RealSRPixRestore27.410.76950.16576.1456.280.31663.0066.74
RealSRSketch2Pix-OS (Ours)25.630.71710.28435.1168.340.51573.58107.01
LSDIRPixRestore21.420.55580.27414.7263.380.32283.6639.81
LSDIRSketch2Pix-OS (Ours)21.150.53990.21394.2570.890.53983.9142.07

References:

[2] Lingchen Sun et al. PixRestore: Unified Image Restoration via Pixel Diffusion Transformer. arXiv preprint arXiv:2608.16793, 2026.

We hope these explanations clear up your confusion. We have provided clearer explanations and experimental analysis in the revised version. If you have any further concerns, we would be pleased to provide more detailed clarification.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer Seub,

Thank you for the careful evaluation of our paper. We want to know whether our previous response has relieved or successfully addressed any of your initial concerns. If not and you have any further questions, please feel free to share them with us so that we can use this last chance to address them. We want to thank you again for your valuable time and efforts. We will be more than happy to receive your response and resolve any further problems.

Best regards,

The Authors

Reviewer_kkAqOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses Real-World Image Super-Resolution (Real-SR) in pixel space, where existing pixel-space diffusion models avoid VAE compression but couple structural reconstruction and detail generation along a single denoising trajectory. The authors propose Sketch2Pix, a self-contained pixel-space diffusion framework that explicitly reorders the denoising process so that structural recovery precedes detail synthesis (“Sketch Before Pixel”).

Critical strengths
  1. The Dual-Time Joint Diffusion and Structure-Forcing Schedule provide a principled way to reorganize the denoising trajectory so that low-frequency structure guides high-frequency detail synthesis. This directly addresses a limitation of single-trajectory pixel-space diffusion and is supported by trajectory analysis and ablations.
  2. Sketch2Pix avoids VAE compression and external pre-trained encoders, performing restoration directly in RGB space with a shared DiT backbone and lightweight DiP decoder. The one-step distilled model achieves a favorable balance between reconstruction fidelity and perceptual quality, with lower parameter count and latency than several alternatives.
Critical weaknesses
  1. The structural target is a low-frequency image obtained by zeroing all DWT detail subbands. While simple, this may discard important mid-frequency structural cues (e.g., thin edges, small text) that are crucial for Real-SR. The fixed α=16 is chosen from a limited exploration; no adaptive or content-aware scheduling is proposed. It is unclear whether the same schedule works for different degradation levels, resolutions, and scene types.
  2. In one-step distillation, the teacher is conditioned on HR-derived structure while the student uses LR-derived structure during both training and inference. The paper does not quantify how much this mismatch limits the student’s performance or whether an alternative conditioning would be better.
Questions for authors
  1. How sensitive is Sketch2Pix to the choice of low-frequency representation? Have you considered retaining more DWT subbands (e.g., LL+LH) or using a learnable structural target? What is the impact on text and fine-edge fidelity?
  2. How does performance vary with α across different datasets and degradation types? Is α=16 universally optimal, or does it require tuning? Could an adaptive schedule based on image content or noise level improve results?
  3. In one-step distillation, the student uses LR-derived structure while the teacher uses HR-derived structure. How much does this mismatch degrade performance compared to a student trained with HR-derived structure (if available at inference)? Can you provide an ablation?
  4. Are there cases where structure-first denoising hurts performance (e.g., highly textured images without clear low-frequency structure)? Can you show qualitative or quantitative failure analysis?
3 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

We would like to sincerely thank the reviewer for the detailed and constructive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.

Weakness 1 & Question 1: Choice of Low-Frequency Structure and Its Impact on Detail Recovery

Answer: Thank you for your insightful comment. We agree that the LL representation does not fully preserve fine edges and small text. However, the structural branch is intended to provide coarse-scale spatial guidance, while the full-image branch restores fine details.

Following your suggestion, we have replaced the LL structural target with LL+LH and conducted a comparison on ScreenSR to assess whether retaining more subbands improves reconstruction. As shown in Table R1, the LL structural target achieves better results on most evaluated metrics. This suggests that retaining the additional LH detail subband does not consistently improve the final reconstruction, providing empirical support for using a compact low-frequency target for spatial guidance. The corresponding results and discussion have been added to the revised manuscript.

Table R1. Comparison of LL and LL+LH structural targets on ScreenSR. Bold indicates the better result.

Structural representationPSNR↑SSIM↑LPIPS↓NIQE↓MUSIQ↑MANIQA↑DeQA↑FID↓
Wavelet (LL+LH)22.890.63320.22573.5469.300.48243.5261.43
Wavelet (LL, Ours)23.680.68170.20963.8071.110.50683.8162.53

Weakness 1 & Question 2: Selection and Cross-Dataset Robustness of the Scheduling Parameter α

Answer: Thank you for your insightful question. In this work, α controls the relative denoising progress of structure and image content. It allows structure to emerge moderately ahead of detail synthesis, providing early guidance while preserving their continued joint refinement. Thus, the schedule is designed to coordinate structural recovery and detail synthesis, rather than to exploit dataset-specific properties.

As shown in Figure 4 and Table 4, α=16, selected on RealSR, also achieves the best performance among the tested values on ScreenSR (512×512 inputs) and LSDIR (128×128 inputs), despite differences in scene content, degradation characteristics, and input resolution. We therefore use α=16 across all evaluated datasets without dataset-specific tuning. These results further support the cross-dataset applicability of this setting.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2 & Question 3: Effect of Student Structural Conditioning

Answer: Thank you for your valuable suggestion. To quantify the effect of student structural conditioning, we have added a controlled comparison in which the student uses either LR- or HR-derived low-frequency structure during both training and testing, with all other settings held constant. The HR condition corresponds to the idealized setting raised by the reviewer, where HR structure is available at inference.

As shown in Table R2, replacing LR conditioning with HR conditioning increases PSNR from 25.63 to 41.14 (a gain of 15.51 dB), reduces LPIPS from 0.2843 to 0.0234, and reduces FID from 107.01 to 5.12. However, these gains are accompanied by lower performance on NIQE, MUSIQ, MANIQA, and DeQA, suggesting that more accurate structural conditioning improves fidelity without necessarily yielding better no-reference perceptual quality.

Notably, using HR structure not only aligns the student’s structural conditioning with the teacher’s but also provides the student with target low-frequency information unavailable at inference. Therefore, the substantial fidelity gains observed in this comparison reflect the overall benefits of ideal structural conditioning and cannot be attributed solely to eliminating the teacher–student conditioning mismatch. In our distillation framework, the teacher uses HR structure to generate reliable supervision anchors, while the student learns to approximate these anchors using LR structure available in practice. This asymmetric conditioning is designed to transfer the teacher’s restoration capability, supported by additional information, to a deployable student while maintaining consistent student conditioning during training and inference. The corresponding results and discussion have been added to the revised manuscript.

Table R2. Effect of student structural conditioning on RealSR. Bold indicates the better result.

Student structural conditioningPSNR↑SSIM↑LPIPS↓NIQE↓MUSIQ↑MANIQA↑DeQA↑FID↓
LR-derived structure (Ours)25.630.71710.28435.1168.340.51573.58107.01
HR-derived structure (Oracle)41.140.97880.02347.131359.450.35353.025.12

Question 4: Failure Cases and Limitations of Structure-First Denoising

Answer: Thank you for your suggestion. In the highly textured examples evaluated, we have not observed systematic quality degradation under the chosen schedule. However, this does not imply that structure-first denoising benefits all types of content. Figure 4 and Table 4 show that increasing structural precedence beyond a moderate level can reduce restoration performance, indicating that earlier structural recovery is not always better. We therefore adopt moderate structural precedence to provide early guidance for detail synthesis while allowing continued joint refinement of structure and image content.

Thank you again for your valuable feedback. We have revised the paper to incorporate all related experiments and detailed discussions as you suggested. If there are any additional issues, we are happy to provide further explanations.

AuthorsOpenReview ↗

Acknowledgment to the reviewer kkAq

Dear Reviewer kkAq,

Thank you for recognizing our writing and ideas and raising fixable concerns. During the review, you provided very constructive comments and perspectives. We feel fortunate to receive patient, meticulous, and responsible reviews.

You have provided constructive comments through careful analysis. Based on your feedback, we have tried our best to address and respond to your concerns. We sincerely appreciate your detailed analysis and hope that our revision has resolved the fixable concerns you raised.

According to your suggestions, we supplemented the experiments, clarified statements, and the experimental results further highlight our strengths. Strictly speaking, this has become a work improved through your efforts.

Careful and time-consuming reviews are not easy. We appreciate your efforts and understand your busy schedule. Regardless of the outcome, the time and effort you have devoted deserve our sincere thanks.

Thank you again for your contribution in refining this work.

Wishing you all the best!

Best regards,

The Authors

Reviewer_vcTyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Sketch2Pix, a VAE-free pixel-space diffusion Transformer for real-world image super-resolution. The method constructs a low-frequency structural target by retaining the LL wavelet subband and applying the inverse transform, and jointly predicts clean structure and the full image using separate diffusion times. A structure-forcing schedule advances structural denoising ahead of image denoising to guide subsequent image refinement. The authors further distill the multi-step model into a one-step student conditioned on LR-derived structure, combining teacher velocity matching, shortcut consistency, and dual-anchor supervision from teacher outputs and HR images. Experiments on RealSR, LSDIR, and ScreenSR, together with component and scheduling ablations, demonstrate competitive restoration quality and efficient one-step inference under the evaluated settings.

Critical strengths
  1. Clear motivation and presentation. The paper has a coherent narrative, and the method is developed around its stated motivation. The clear and accessible presentation makes the proposed framework and its design choices easy to follow.

  2. Strong restoration results with efficient one-step inference. The proposed method performs well across multiple benchmarks. In particular, the one-step model achieves competitive restoration quality with low measured inference latency under the reported experimental setting, supporting its practical potential.

Critical weaknesses
  1. Limited conceptual novelty relative to Latent Forcing. The central motivation and the analysis of denoising order closely resemble those in Latent Forcing. The current presentation does not sufficiently establish the conceptual contribution beyond using a wavelet-defined low-frequency structural representation and adapting the framework to real-world super-resolution.

  2. Insufficient detail about one-step distillation. The manuscript does not specify the procedure used to generate the teacher anchor images. This omission makes it difficult to assess whether the teacher is used consistently with its training trajectory and to reproduce the distillation procedure.

Questions for authors
  1. Relationship to Latent Forcing. Jointly modeling auxiliary representations and pixels with separate noise times so that the auxiliary representation is recovered earlier is already central to Latent Forcing. Replacing the auxiliary representation with wavelet-derived low-frequency structure appears to be a relatively incremental change. Could the authors clarify the specific differences in motivation and methodology, and explain what new insight is established for real-world super-resolution beyond the existing ordering principle?

  2. Comparison with pixel-space methods. The main comparisons do not include external pixel-space restoration models. Could the authors provide a comparison with PixRestore under a matched super-resolution evaluation protocol? This would help establish the proposed method's advantages over relevant pixel-space alternatives, rather than only the latent-space baselines currently evaluated.

  3. Teacher anchor generation and distillation ablations. The multi-step model starts inference at (t,τ)=(1,1)(t,\tau)=(1,1) and follows the trajectory used during training. However, the distillation procedure describes a teacher conditioned on clean HR-derived structure. Does teacher anchor generation start at (t,τ)=(1,0)(t,\tau)=(1,0), and, if so, does this move the teacher outside its training trajectory? Please specify the complete procedure for generating x^T\hat{x}^{T}, including the time sequence and how the structural state is handled. In addition, the distillation objective contains three terms. Could the authors provide ablations showing the individual contributions of velocity matching, shortcut consistency, and dual-anchor supervision?

  4. Scalability to higher resolutions. The main experiments and efficiency evaluations use 512×512512\times512 outputs. Could the authors evaluate 4× super-resolution at substantially higher output resolutions, preferably around 4K (e.g., 960×540960\times540 inputs to 3840×21603840\times2160 outputs)? Please clarify any input-size constraints and whether padding or tiled inference is required, and report restoration quality, end-to-end latency, and peak GPU memory against representative pixel-space and latent-space baselines under the same hardware setting.

5 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

We would like to sincerely thank the reviewer for the detailed and constructive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.

Weakness 1 & Question 1: Relationship to Latent Forcing

Answer: Thank you for your constructive feedback. We would like to clarify the contributions of our work from three perspectives: Motivation, Method, and New Insight for Real-world Image Super-Resolution (Real-SR).

(1) Motivation. Latent Forcing [1] mainly investigates how auxiliary representations facilitate pixel generation, while we focus on how to recover reliable structural constraints in Real-SR from incomplete and distorted structural cues, enabling detail reconstruction that balances structural fidelity and visual realism. To this end, we treat structure as an explicit recovery objective that continuously updates during joint denoising, providing progressive guidance for detail reconstruction.

(2) Method. We adopt a computationally efficient, learning-free wavelet transform to define a spatially aligned, coarse-scale structural target. Conditioned on the LR input, we jointly restore the structure and the full image while independently controlling their denoising progress. Unlike the pretrained configurations in Latent Forcing, our design requires no external encoder and forms a self-contained, end-to-end pixel-space framework for Real-SR. Meanwhile, to reduce inference cost, we further distill the multi-step model into a one-step student for efficient restoration.

(3) New Insight for Real-SR. We find that jointly restoring structure and coordinating its progress with detail synthesis can improve Real-SR reconstruction compared with conditioning on fixed structural cues from degraded LR images. Specifically, Figure 6 demonstrates the benefits of joint structural recovery over fixed LR structural conditioning. Figure 4 and Table 4 further show that moderate structural precedence improves reconstruction, whereas excessive precedence degrades quality. Motivated by these findings, we propose the Sketch Before Pixel paradigm, which explicitly coordinates structural recovery and detail synthesis within joint denoising to balance structural fidelity and visual realism.

Question 2: Comparison with Pixel-Space Methods

Answer: Thank you for your helpful suggestion. As suggested by the reviewer, we have added quantitative and qualitative comparisons with PixRestore [2] under a matched evaluation protocol.

(1) Quantitatively, PixRestore achieves higher PSNR and SSIM but substantially lower perceptual quality scores than our method, as shown in Table R1. Specifically, on RealSR and LSDIR, our method exceeds PixRestore by 12.06 and 7.51 points in MUSIQ and 0.1991 and 0.2170 in MANIQA, respectively. We exclude ScreenSR to avoid potential training–test overlap arising from PixRestore’s different data split.

(2) Qualitatively, the examples in Figure 5 of the revised manuscript further demonstrate that our method more faithfully preserves fine textures and text structures while recovering more realistic details than PixRestore.

The complete comparison results and accompanying discussion have been added to the revised manuscript.

Table R1. Quantitative comparison with PixRestore on LSDIR and RealSR. Bold indicates the better result.

DatasetMethodPSNR↑SSIM↑LPIPS↓NIQE↓MUSIQ↑MANIQA↑DeQA↑FID↓
LSDIRPixRestore21.420.55580.27414.7263.380.32283.6639.81
LSDIRSketch2Pix-OS (Ours)21.150.53990.21394.2570.890.53983.9142.07
RealSRPixRestore27.410.76950.16576.1456.280.31663.0066.74
RealSRSketch2Pix-OS (Ours)25.630.71710.28435.1168.340.51573.58107.01

References:

[1] Alan Baade et al. Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation. International Conference on Machine Learning (ICML), 2026.

[2] Lingchen Sun et al. PixRestore: Unified Image Restoration via Pixel Diffusion Transformer. arXiv preprint arXiv:2608.16793, 2026.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2 & Question 3: Teacher Anchor Generation and Distillation Ablations

Answer: Thank you for your valuable question and suggestion. We would like to clarify the following two aspects in detail.

(1) Teacher Anchor Generation. Since one-step inference cannot progressively recover structure before synthesizing details, we provide structural conditioning at the start of prediction to retain its guidance for detail reconstruction. The teacher uses clean HR low-frequency structure as a reliable spatial reference, whereas the student uses LR low-frequency structure available at inference. To generate teacher anchors, we initialize the image with Gaussian noise and perform 25 Euler updates from (t,τ)=(1,0)(t,\tau)=(1,0) along tk=1−k/25t_k=1-k/25, updating only the image while keeping the HR structure and τ=0\tau=0 fixed. The final output is cached as the teacher anchor, and the student makes a one-step prediction from the same initial image noise. Although this sampling path differs from the original joint training trajectory, it provides fixed, reliable structural guidance for anchor generation. The one-step restoration results support the effectiveness of these anchors as distillation supervision.

(2) Distillation Loss Ablations. Following your suggestion, we have added ablation experiments to assess the contributions of velocity matching, shortcut consistency, and dual-anchor supervision, as reported in Table R2. Removing velocity matching causes severe reconstruction degradation, while removing shortcut consistency worsens PSNR, SSIM, and four other quality metrics, despite slight improvements in LPIPS and NIQE. In addition, removing dual-anchor supervision worsens all reported metrics, highlighting its contribution to both reconstruction fidelity and perceptual quality. These results support the complementary roles of the three losses in constraining the denoising direction, promoting one-step consistency, and supervising reconstruction quality.

The complete updated table and accompanying discussion have been added to the revised manuscript.

Table R2. Ablation study of one-step distillation losses on RealSR. Bold indicates the best result.

VariantPSNR↑SSIM↑LPIPS↓NIQE↓MUSIQ↑MANIQA↑DeQA↑FID↓
w/o Velocity Matching14.840.19390.617223.9644.820.36751.52325.19
w/o Shortcut Consistency25.080.69810.28375.0366.550.44623.42107.85
w/o Dual-Anchor Supervision23.590.68650.45049.3225.580.22531.78141.05
Full model (Ours)25.630.71710.28435.1168.340.51573.58107.01
AuthorsOpenReview ↗

Official Response by Authors – Part 3

Question 4: Scalability to Higher Resolutions

Answer: Thank you for your valuable suggestion. We have added a 4× super-resolution evaluation on UHD4K, reconstructing 3840×21603840\times2160 outputs from 960×540960\times540 inputs. We compare Sketch2Pix-MS and Sketch2Pix-OS with representative pixel-space and latent-space baselines on a single NVIDIA H800 GPU. Our method processes 512×512512\times512 output-space tiles with a 64-pixel overlap and combines them through weighted blending, enabling full-resolution reconstruction with bounded memory usage per tile.

Table R3 reports restoration quality, end-to-end latency, and peak allocated GPU memory. It can be seen that Sketch2Pix-MS and Sketch2Pix-OS achieve the highest MUSIQ and MANIQA scores, respectively, with the one-step variant requiring 10.65 s per image and 3.25 GiB of peak allocated GPU memory. Although PixRestore achieves higher PSNR and SSIM, its substantially lower MUSIQ and MANIQA scores reflect a trade-off between fidelity and perceptual quality. These results demonstrate scalability to 4K reconstruction and the efficiency gains of our one-step variant over the multi-step version. The corresponding experiments and discussion have been added to the revised manuscript.

Table R3. Reconstruction quality and computational efficiency for 4× super-resolution on UHD4K. Bold indicates the best result within each inference setting.

MethodPSNR ↑SSIM ↑MUSIQ ↑MANIQA ↑Latency ↓ (s/image)Peak GPU Memory ↓ (GiB)
StableSR24.750.789834.020.26781123.4012.36
DiT4SR21.420.681649.040.4609275.2347.85
DiT-SR25.890.790642.160.380557.6520.76
VOSR22.580.780348.320.428196.0126.52
Sketch2Pix-MS (Ours)26.130.804354.220.496670.326.11
SinSR25.610.780542.730.39446.933.57
OSEDiff22.520.744953.130.452852.179.17
CODSR22.210.716153.710.475058.319.39
PixRestore26.560.807240.310.29767.533.21
Sketch2Pix-OS (Ours)26.220.792453.830.516010.653.25

Thank you again for your valuable feedback. We have revised the paper to incorporate all related experiments and detailed discussions as you suggested. If you have any further concerns, we would be pleased to provide more detailed clarification.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair

Dear Area Chairs,

We sincerely thank the reviewers and area chairs for their valuable time and constructive feedback.

Following our detailed rebuttal, Reviewer vcTy raised their rating to 3 (Weak Acceptance), and Reviewer kkAq’s current rating is also positive at 3 (Weak Acceptance). Reviewer Seub acknowledged that our rebuttal resolved their concerns about the experimental settings and that the experiments and writing had improved and were sound. Below, we provide a brief summary of the author–reviewer discussion.

  • Reviewer vcTy recognized our “Clear motivation and presentation” and “Strong restoration results with efficient one-step inference.” In our rebuttal, we clarified the relationship to Latent Forcing and the teacher anchor generation procedure, added comparisons with PixRestore and ablations of the distillation losses, and evaluated 4K super-resolution quality and efficiency. The reviewer stated that these additions “provide convincing evidence of the method’s effectiveness” and raised their rating to 3 (Weak Acceptance).

  • Reviewer kkAq recognized that our Dual-Time Joint Diffusion and Structure-Forcing Schedule “provide a principled way to reorganize the denoising trajectory” and that the one-step model achieves “a favorable balance between reconstruction fidelity and perceptual quality.” In response, we added an LL versus LL+LH structural-target comparison and an ablation of LR- versus HR-derived student conditioning, clarified the schedule’s cross-dataset applicability, and discussed the limitations of excessive structural precedence. The reviewer assigned a positive rating of 3 (Weak Acceptance), recognizing the framework’s design and restoration performance.

  • Reviewer Seub described “Sketch Before Pixel” as “intuitive and telling” and confirmed that our rebuttal resolved their concerns about experimental settings, acknowledging that “The experiment and writing are improved and sound.” To address the remaining novelty concern, we further clarified the contribution of jointly restoring structure and image content with controlled recovery order, supported by experiments showing the benefits of moderate structural precedence. We also added LOL-v1 results, where Sketch2Pix achieves the highest PSNR and SSIM among the compared methods alongside competitive LPIPS, providing further evidence of its restoration capability and potential beyond super-resolution.

In this paper, we propose Sketch2Pix, a self-contained pixel-space restoration framework that jointly recovers structure and image content without a VAE or an external pretrained encoder. Its central insight is that moderate structural precedence can improve detail reconstruction, making recovery order an explicit design choice for Real-World Image Super-Resolution. Together with efficient one-step distillation, the results demonstrate competitive restoration quality, scalability to 4K outputs through tiled inference, and promising applicability to low-light enhancement.

We hope that the additional evidence and clarifications will support your consideration of our work. Once again, we sincerely thank the reviewers and area chairs for their valuable time and constructive feedback, which have helped us strengthen the paper.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.