acceptodds
Under review as a conference paper at ICLR 2027

PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution

10 views

Abstract

Latent diffusion models have recently achieved remarkable progress in Real-world Image Super-Resolution (Real-SR). However, existing methods consistently exhibit a persistent seesaw effect: enhancing perceptual realism compromises structural fidelity. This effect, we argue, is not intrinsic to diffusion models, but largely stems from the architectural constraints of prevailing two-stage frameworks. To address this challenge, we propose PixelSR, a novel single-stage Pixel Diffusion Transformer for Real-SR that operates directly in pixel space. Specifically, PixelSR consists of two core designs: Pixel-space Conditional Diffusion and Orthogonal Flow Decoupling. The former establishes an effective pixel-space restoration framework by integrating spatially aligned LR conditioning, visual semantic guidance, and lightweight texture refinement, thereby bypassing VAE compression and enabling end-to-end optimization directly in RGB space. Complementarily, the latter decomposes the asymmetric flow target into a low-rank semantic component and a noise-free orthogonal texture component. By routing these components to specialized branches, it achieves decoupled optimization with reduced mutual interference. Benefiting from these designs, our PixelSR restores perceptually realistic details while faithfully preserving structural content. Extensive experiments across multiple benchmarks demonstrate that PixelSR not only achieves competitive performance, but also effectively mitigates the fidelity–realism seesaw effect.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pdV1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a technically interesting pixel-space Real-SR framework with an exact algebraic decomposition of the asymmetric flow target and encouraging results across LSDIR, RealSR, ScreenSR, and the newly introduced UGCSR benchmark.

Critical strengths

PixelSR presents a coherent pixel-space Real-SR framework with a particularly well-controlled evaluation of Orthogonal Flow Decoupling: the comparison against monolithic asymmetric-flow prediction uses the same network, parameter count, and training setup, making the observed gains convincing. The paper is also clearly presented and demonstrates consistent fidelity–perception performance across several benchmarks, with useful analyses of branch specialization.

Critical weaknesses

The novelty of the individual ingredients is somewhat incremental: the orthogonal decomposition and Procrustes-aligned subspace are already explicit in AsymFlow, while the global-DiT/local-pixel-refiner design is closely related to PixelDiT, so the main novelty is primarily the specialized branch assignment for conditional Real-SR. In addition, the claim that the fidelity–realism seesaw largely originates from the latent two-stage architecture is stronger than the controlled evidence, and the realism claim would benefit from human perceptual evaluation rather than relying mainly on no-reference IQA metrics.

Questions for authors
  1. The orthogonal decomposition in Eq. (14) is algebraically valid, but orthogonality of prediction errors in output space does not by itself imply non-conflicting gradients in parameter space, particularly because the texture branch is conditioned on DiT features. The statement that OFD enables joint training “without mixing optimization signals” is therefore stronger than what is currently demonstrated. Please report per-step cosine similarities between the semantic- and texture-objective gradients on shared parameters and compare them against the monolithic u_A baseline to directly validate the reduced-interference claim.

  2. The manuscript attributes the seesaw effect largely to VAE compression and the two-stage latent pipeline, but comparisons against independently trained latent-SR systems cannot establish this causal explanation because architecture, parameterization, training data, model scale, and optimization all vary simultaneously. Please add a tightly controlled pixel-versus-latent experiment using matched backbone capacity, training pairs, optimization budget, and conditioning, together with the reconstruction ceiling imposed by the corresponding VAE. This would separate the benefit of removing the VAE from the benefits of PCD, x_0-prediction, semantic conditioning, and OFD.

  3. PixelDiT already combines patch-level global modeling with pixel-level detail modeling, DiP similarly decouples global construction and local detail restoration, and AsymFlow introduces the low-rank asymmetric velocity parameterization from which OFD is derived. Please add a component-comparison table covering prediction target, global/local branches, low-rank projection, projector construction, conditioning mechanism, parameter count, FLOPs, and training/inference cost, and clearly identify which capability is uniquely introduced by OFD rather than inherited from these ingredients.

  4. The training hardware, total training compute, and comparable resource statistics of the baselines are unspecified. Please report parameters, MACs/FLOPs, peak inference memory, sampling steps, training data scale, training hardware/time where available, and a compute-versus-quality Pareto curve in addition to raw wall-clock inference time.

  5. Please release or precisely document the benchmark construction pipeline, evaluation splits, model/configuration files, baseline versions, and all preprocessing necessary to reproduce Tables 1–6.

  6. An adversarial threat model is not applicable to this restoration paper; the analogous experimental contract is the assumed degradation and deployment model, which should be stated more explicitly, including scale factor, spatial alignment, degradation family, and access to the frozen semantic encoder. The semantic-versus-texture interpretation of P and I-P is currently supported mainly by visualizations, spectra, and t-SNE, so stronger tests under unseen degradations, mild LR–HR misregistration, and alternative scale factors would help establish that the decomposition reflects a reusable property rather than the specific training pipeline. Please also provide a dense sweep of the orthogonal-branch scale \alpha and plot PSNR/SSIM/LPIPS and perceptual-IQA metrics jointly to verify whether the claimed fidelity–realism behavior forms a stable Pareto trajectory.

7 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.

Question 1: Effectiveness and Gradient Interference Analysis of Task-Aligned Branch Assignment

Answer: Thank you for your constructive feedback. Indeed, orthogonality in the output space does not necessarily imply conflict-free gradients on shared parameters, particularly because PiT depends on DiT features. Accordingly, “decoupling” in our work refers to separating the complementary prediction targets, rather than achieving complete gradient independence on shared parameters.

Following your suggestion, we have compared the gradient relationships on shared parameters across OFD, target swapping, and the undecomposed uAu_A baseline. As shown in Table R1, OFD reduces the negative-cosine rate from 5.75% to 3.85% and the mean negative magnitude from 0.0076 to 0.0033 relative to the undecomposed baseline. Both measures are also lower than those of the swapped-assignment variant. These results provide evidence of reduced negative gradient conflict in the observed records, complementing the restoration-quality comparisons. We have revised the phrase “without mixing optimization signals” to avoid equating target separation with conflict-free optimization.

Table R1: Effectiveness Analysis of Task-Aligned Branch Assignment. Bold indicates the best result.

MethodPSNR ↑SSIM ↑LPIPS ↓NIQE ↓MUSIQ ↑MANIQA ↑DeQA ↑FID ↓Mean Gradient CosineNegative Cosine Rate (%) ↓Mean Negative Magnitude (×10⁻³) ↓
uAu_{\mathrm A}24.790.68450.31575.3869.570.46803.55117.920.22015.757.60
Target Swapping24.880.69270.31614.7169.410.51803.49118.150.169210.578.50
OFD (Ours)25.340.71660.29225.5769.590.53813.61112.890.24183.853.30

Question 2: Controlled Pixel–Latent Comparisons and VAE Reconstruction Analysis

Answer: Thank you for your valuable feedback. To distinguish the contribution of the modeling space from those of other designs, Figure 5 compares latent-space and pixel-space DiT under a unified training protocol, keeping the DiT configuration and vv-prediction unchanged, before progressively introducing x0x_0-prediction, semantic conditioning, PiT, and OFD. The results show that directly moving to pixel space does not improve performance, whereas the subsequent designs yield further gains. This indicates that removing the VAE alone is insufficient to explain PixelSR’s performance advantage, and that prediction and supervision designs tailored to pixel-space modeling are also important.

We have additionally conducted deterministic HR reconstruction experiments using the VAE’s native encoder and decoder, with the posterior mean, to separately examine the distortion introduced by encoding and decoding. As shown in Table R2, the reconstructions exhibit measurable deviations from the original HR images across all four datasets, further confirming that VAE encoding and decoding introduce reconstruction errors.

Table R2: Quality Comparison between VAE Reconstructions and Original HR Images. Each entry is reported as VAE reconstruction / original HR.

DatasetPSNR ↑SSIM ↑LPIPS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑DeQA ↑
LSDIR24.68 / ∞\infty0.7168 / 1.000.0904 / 0.0018.59 / ≈03.74 / 3.6672.52 / 72.870.5399 / 0.56054.07 / 4.06
RealSR35.31 / ∞\infty0.9210 / 1.000.0364 / 0.0013.37 / ≈06.77 / 6.1257.34 / 57.460.3317 / 0.32873.07 / 3.05
ScreenSR25.55 / ∞\infty0.7645 / 1.000.0812 / 0.0023.98 / ≈03.87 / 3.7771.77 / 72.150.5434 / 0.56504.07 / 4.12
UGCSR27.86 / ∞\infty0.7889 / 1.000.0654 / 0.0024.07 / ≈04.00 / 3.8974.88 / 75.170.5767 / 0.58834.15 / 4.18
AuthorsOpenReview ↗

Official Response by Authors – Part 2

Question 3: Component Comparison and OFD Contributions

Answer: Thanks for your valuable suggestion. We would like to clarify that our work adopts only the PiT architecture from PixelDiT and the asymmetric prediction target and low-rank projection from AsymFlow. Building on these components, our PCD introduces pixel-space conditioning tailored to Real-SR, while OFD assigns the complementary projected targets to the globally modeling DiT and locally modeling PiT branches. The two branches thus predict semantic and texture components, respectively, establishing explicit task specialization.

Following your suggestion, we have added a component comparison in Table R3 to distinguish inherited designs from PixelSR’s contributions, alongside parameter counts and training-data scales. PixelDiT, DiP, and AsymFlow are evaluated on image generation, with different tasks, resolutions, and inference settings from PixelSR, so inference times are not directly comparable.

In addition, the target-swapping ablation in Table R1 keeps the architecture, target decomposition, and training settings fixed while exchanging only the prediction targets assigned to the two branches. Our OFD outperforms the swapped variant, further supporting the effectiveness of matching prediction targets to the modeling characteristics of each branch.

Table R3: Component Comparison of PixelSR and Related Methods.

MethodTaskPrediction targetGlobal/local modelingLow-rank projectionLR + DINOv2 guidanceOFDParametersTraining data
PixelDiT-XL/16Image generationϵ−x0\epsilon-x_0✓✗✗✗797MImageNet
DiP-XL/16Image generationϵ−x0\epsilon-x_0✓✗✗✗631MImageNet
AsymFlow-H/16Image generationuA=Pϵ−x0u_A=P\epsilon-x_0✗✓✗✗953MImageNet
PixelSR (Ours)Real-SRPuAPu_A to DiT; (I−P)uA(I-P)u_A to PiT✓✓✓✓773.4MLSDIR, DIV2K, DIV8K, OST, Flickr2K, 10K FFHQ, and 50K web images

Question 4: Computational Cost and Efficiency

Answer: Thanks for your helpful suggestion. We have added parameters, NFE, and measured inference time in Table R4, alongside PSNR and MUSIQ on UGCSR to compare restoration quality and efficiency. Runtime is measured for 512×512512\times512 outputs on an H800 GPU with a batch size of 1. NFE denotes the number of model evaluations.

Table R4: Quality and Efficiency Comparison on UGCSR. Bold indicates the best result, and italics indicate the second-best result.

MethodNFEPSNR ↑MUSIQ ↑Time (s) ↓Params (M) ↓
StableSR20024.4070.825.711540
DiffBIR5023.0874.015.19380†
DiT4SR4021.7472.945.1721380
VOSR2521.8974.743.86619.4
DiT-SR1523.9669.931.8761
OSEDiff122.4073.580.461760
SinSR123.8269.270.33174
CODSR122.2374.986.2135.79
PixelSR (Ours)2525.7275.694.44773.4
AuthorsOpenReview ↗

Official Response by Authors – Part 3

Question 5: Reproducibility Details and Release Plans

Answer: Thank you for your valuable feedback. We train on LSDIR, DIV2K, DIV8K, OutdoorSceneTraining, Flickr2K, 10K FFHQ images, and 50K web images, with LR–HR pairs synthesized using the Real-ESRGAN pipeline. For synthetic evaluation, we center-crop LSDIR validation images to 512×512512\times512 and generate 128×128128\times128 LR inputs. Real-world evaluation uses RealSR, ScreenSR, and UGCSR with spatially corresponding crops. Table 4 in the supplementary material provides the main training and model configurations.

We also plan to release the benchmark, model configurations, and training and evaluation scripts upon publication to support reproducibility.

Question 6: More Details and Validation

Answer: Thank you for the suggestion. We provide further details and analysis in the following three aspects.

(1) Implementation details. We train the model for 4×4\times super-resolution using spatially aligned LR–HR image pairs, with resolutions of 128×128128\times128 and 512×512512\times512, respectively. Training uses synthetic degradations, while evaluation covers camera-captured images, screen content, and platform-distributed images. DINOv2 remains frozen during both training and inference and extracts semantic conditioning solely from the LR input, without requiring HR images or additional annotations for this conditioning.

(2) Evaluation across restoration settings. We evaluate on RealSR, ScreenSR, and UGCSR, covering different scenarios and their associated degradations. RealSR uses a 4×4\times super-resolution setting, whereas ScreenSR and UGCSR use 1×1\times same-resolution restoration settings. These results across scenarios and scale settings complement the branch ablations and support the applicability of the decomposition to different restoration settings.

(3) Fidelity–realism trade-off. We evaluate α∈{0,0.5,1,1.5}\alpha\in\{0,0.5,1,1.5\} on RealSR by varying α\alpha only at the final sampling step, while keeping α=1\alpha=1 for the preceding 24 steps. Table R5 reports PSNR, SSIM, LPIPS, MUSIQ, and MANIQA. Increasing α\alpha from 0 to 1 improves all five metrics. Further increasing it to 1.5 yields higher MUSIQ and MANIQA but lowers PSNR and SSIM and increases LPIPS. These results support α=1\alpha=1 as the default setting and show that the final orthogonal-branch update can influence the fidelity–perceptual quality trade-off.

Table R5: Effect of the Orthogonal-Branch Scale α\alpha on RealSR. Bold indicates the best result.

α\alphaPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑MANIQA ↑
021.930.61430.334559.440.3715
0.524.890.70140.308365.840.4533
125.340.71660.292269.590.5381
1.523.260.63270.331872.670.5567

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.

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

The paper propose PixelSR, a single-stage Pixel Diffusion Transformer for Real-SR that operates directly in pixel space. Specifically, PixelSR consists of two core designs: Pixel-space Conditional Diffusion and Orthogonal Flow Decoupling. The former establishes an effective pixel-space restoration framework by integrating spatially aligned LR conditioning, visual semantic guidance, and lightweight texture refinement, thereby bypassing VAE compression and enabling end-to-end optimization directly in RGB space. Complementarily, the latter decomposes the asymmetric flow target into a low-rank semantic component and a noise-free orthogonal texture component. By routing these components to specialized branches, it achieves decoupled optimization with reduced mutual interference. Benefiting from these designs, PixelSR restores perceptually realistic details while faithfully preserving structural content. Extensive experiments across multiple benchmarks demonstrate that PixelSR not only achieves competitive performance, but also effectively mitigates the fidelity–realism seesaw effect.

Critical strengths

1.The paper addresses the persistent trade-off between perceptual realism and structural fidelity in real-world image super-resolution. This challenge is particularly relevant to diffusion-based SR methods, which may generate visually plausible details that are inconsistent with the underlying image structure. The problem is well motivated, clearly articulated, and closely related to practical challenges in restoring severely degraded images. 2.The proposed pixel-space conditional flow formulation offers a reasonable alternative to conventional latent-space restoration pipelines. By integrating spatially aligned low-resolution conditioning, frozen DINOv2 semantic features, and a lightweight pixel-level decoder, the method combines structural guidance, high-level semantic information, and texture refinement within a unified restoration framework. 3.The OFD module decomposes the asymmetric velocity target into two complementary components intended to capture different aspects of the restoration process, which are then handled by specialized network branches. This design provides a clear rationale for separating global generation from local reconstruction. The paper further supports its design choices with ablation studies on projector construction, semantic encoder selection, and the scaling of the orthogonal component.

Critical weaknesses

1.Eq. (14) follows directly from applying orthogonal projectors to AsymFlow's asymmetric velocity target. The main additional design is assigning the resulting components to DiT and PiT branches. Although Table 2 reports improvements under matched architecture and parameter count, target factorization and branch assignment are introduced simultaneously, leaving their individual contributions and the substantive novelty unclear. 2. Section 3.1 identifies the 1/t conversion as ill-conditioned, yet Eq. (16) retains this factor for the orthogonal component, weighting its squared prediction error by 1/t^2 in the velocity-space loss. This component spans 752 of 768 dimensions per patch. Appendix A.1 reports a minimum noise scale of 0.05 without sufficiently explaining its application or impact on numerical stability. 3. Table 1 compares single operating points under default inference settings and summarizes results through aggregate rankings. On ScreenSR, PixelSR achieves lower PSNR than DiT-SR and lower MANIQA and DeQA than DiffBIR. These results demonstrate competitive performance but do not establish an improved attainable Pareto frontier without matched-distortion comparisons or trade-off curves. 4. PixelSR reports particularly strong gains on UGCSR, but Appendix A.2 leaves source platforms, processing configurations, alignment procedures, and potential overlap with the 50K web-collected training images insufficiently documented. The appendix also acknowledges that UGCSR contains relatively high-quality LR inputs with subtle degradations. These limitations affect reproducibility and weaken the generalizability of the reported gains to severely degraded images. 5. Poor figure readability and inconsistent UGCSR category descriptions. Several figures, particularly Figures 2, 5, 6, 7, and 8, contain very small labels and annotations that are difficult to read at normal page size. Increasing font sizes would improve readability. Additionally, the caption of Figure 7 states that the UGCSR dataset contains ten semantic categories, whereas the corresponding appendix text describes nine. This inconsistency should be clarified.

Questions for authors
  1. Table 1 demonstrates PixelSR's competitive overall performance across multiple metrics, but the improvements are not consistent across individual metrics, and the average ranking alone does not establish that the fidelity—realism trade-off has been alleviated. Could the authors provide further experiments or analyses showing whether PixelSR improves perceptual realism at comparable fidelity levels, or achieves higher fidelity at comparable realism levels, to more directly support this central claim?
  2. How is the minimum noise scale of 0.05 applied to timestep sampling, the 1/t conversion in Eq. (16), and the inference endpoint? How sensitive are training stability and restoration quality to this value?
  3. Does the observed specialization arise from OFD or from the different architectures and conditioning of DiT and PiT? Could you provide a swapped-assignment ablation and quantitative semantic probes to disentangle these effects?
  4. Which platforms and processing configurations were used to construct UGCSR, and how were scaling, spatial alignment, and sample selection handled? Will the benchmark and its construction protocol be released?
  5. What checks were performed to exclude overlap or near-duplicates between UGCSR and the 50K web-collected training images? Given UGCSR’s relatively mild degradations, how does PixelSR’s advantage vary across degradation severity and content categories?
7 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 3: Contributions of Task-Aligned Branch Assignment

Answer: Thank you for your insightful comment. We would like to clarify that our method builds on AsymFlow’s asymmetric velocity target and orthogonal decomposition. Our key insight is to assign the resulting complementary prediction targets to DiT and PiT according to their global and local modeling strengths, respectively. This assignment establishes a task-aligned paradigm for pixel-space Real-SR, using explicit semantic and texture supervision to alleviate the trade-off between structural fidelity and visual realism.

As suggested, we have added a swapped-assignment ablation that preserves the target decomposition, branch architectures, conditioning inputs, and training settings while exchanging only the prediction targets assigned to the two branches. As shown in Table R1, OFD outperforms the swapped variant across multiple restoration-quality metrics, supporting the effectiveness of matching prediction targets to branch modeling characteristics. Furthermore, Figure 9 shows clearer category-wise clustering in the semantic-branch features and a more dispersed distribution in the texture-branch features, providing complementary qualitative evidence for their distinct roles.

Table R1: Effectiveness Analysis of Task-Aligned Branch Assignment. Bold indicates the best result.

MethodPSNR ↑SSIM ↑LPIPS ↓NIQE ↓MUSIQ ↑MANIQA ↑DeQA ↑FID ↓Mean Gradient CosineNegative Cosine Rate (%) ↓Mean Negative Magnitude (×10⁻³) ↓
uAu_{\mathrm A}24.790.68450.31575.3869.570.46803.55117.920.22015.757.60
Target Swapping24.880.69270.31614.7169.410.51803.49118.150.169210.578.50
OFD (Ours)25.340.71660.29225.5769.590.53813.61112.890.24183.853.30
AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2 & Question 2: Noise-Scale Implementation and Sensitivity Analysis

Answer: Thank you for your helpful suggestion. We follow JiT’s denominator-clipping strategy [1], implementing the 1/t1/t factor in Eq. (16) as 1/max⁡(t,τ)1/\max(t,\tau), with τ=0.05\tau=0.05 by default, to limit error amplification near the clean endpoint. This threshold only clips the denominator in the velocity conversion, while training timesteps remain uniformly sampled from [0,1][0,1]. During inference, we apply the same clipping while integrating from t=1t=1 to t=0t=0, rather than terminating at t=0.05t=0.05. Thus, 0.050.05 is a denominator-clipping threshold, not the minimum training or inference timestep.

As suggested, we have added a sensitivity analysis comparison for τ∈{0.025,0.05,0.1}\tau\in\{0.025,0.05,0.1\}. As shown in Table R2, the default threshold τ=0.05\tau=0.05 achieves the best results across all reported metrics, supporting its selection. We have included these implementation details and the restoration-quality sensitivity analysis in the revised manuscript.

Table R2: Sensitivity Analysis of the Denominator-Clipping Threshold. Bold indicates the best result.

MethodPSNR ↑SSIM ↑LPIPS ↓NIQE ↓MUSIQ ↑MANIQA ↑DeQA ↑FID ↓
τ=0.1\tau=0.123.51920.70740.22674.633970.32140.47493.637170.2646
τ=0.025\tau=0.02523.55720.71060.22334.433670.42930.47223.669269.4702
τ=0.05\tau=0.05 (Ours)23.56400.71560.20824.328472.86130.48243.829469.1307

Weakness 3 & Question 1: More Comprehensive Analysis of the Fidelity–Realism Trade-off

Answer: Thank you for your valuable comment. As suggested, we have added PSNR–DeQA and SSIM–MUSIQ comparisons on ScreenSR to more directly illustrate the trade-off between fidelity and perceptual quality.

As shown in Figure 10, when compared at similar PSNR levels, PixelSR achieves a higher DeQA score than DiT-SR (3.8294 vs. 3.5160), with PSNR values of 23.5640 and 23.6599 dB, respectively. In addition, at similar SSIM levels, PixelSR achieves a higher MUSIQ score than StableSR (72.8613 vs. 63.9757), with SSIM values of 0.7156 and 0.7150, respectively. These comparisons indicate that our PixelSR can improve perceptual quality at comparable fidelity. In addition, the visual comparisons of local regions in Figures 11 and 12 of the supplementary material illustrate its structure preservation and natural texture restoration, complementing the quantitative results. We have incorporated these analyses into the revised manuscript to provide a more comprehensive assessment of the fidelity–realism trade-off.

References

[1] Tianhong Li et al. Back to Basics: Let Denoising Generative Models Denoise. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 36115–36125, 2026.

AuthorsOpenReview ↗

Official Response by Authors – Part 3

Weakness 4 & Question 4 & Question 5: More Details on the UGCSR Benchmark

Answer: Thank you for the suggestion. We have provided more details on UGCSR’s data sources, construction protocol, and evaluation scope.

(1) Data sources and independence. UGCSR contains 230 LR–HR test pairs constructed through the Kuaishou short-video platform and is independent of the 50,000 training images. Specifically, we use file hashes to detect exact duplicates and perceptual hashes to identify potential near-duplicates resulting from resizing or compression, followed by manual verification to exclude overlap between the training and test sets.

(2) Construction and reproducibility. We upload high-quality images to Kuaishou with the high-definition option enabled and download the processed images as LR inputs, retaining the originals as HR references. Degradations arise from actual platform processing, including resizing, compression, and transcoding. We resize the processed images to match the HR references and extract spatially corresponding center crops. We select diverse content across nine categories and manually exclude distorted references and incorrect pairs. We will release the benchmark and construction protocol upon publication.

(3) Performance and scope. UGCSR focuses on degradations associated with real-world platform processing. Even when LR inputs have relatively high overall quality, local textures, edges, and text structures may remain impaired. This setting therefore evaluates not only detail recovery but also the preservation of reliable content and the avoidance of excessive hallucination. PixelSR’s results on UGCSR demonstrate its ability to balance content preservation and detail recovery, complementing the evaluations on RealSR and ScreenSR.

Weakness 5: Figure Readability and Category Consistency

Answer: Thank you for your careful review. We have increased the font sizes of labels and annotations in Figures 2, 5, 6, 7, and 8 to improve readability at standard page size. We have also rechecked the UGCSR category definitions and aligned the category counts and descriptions in the Figure 7 caption with those in the appendix.

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.

Reviewer_PZWQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PixelSR for real-world image super-resolution in pixel space, combining low-quality image conditioning, DINO semantic features, and a DiT/PiT architecture. Its core method, OFD, assigns complementary projection components of an asymmetric flow target to two branches to balance structural fidelity and texture generation. Its effectiveness is evaluated across multiple datasets and through component ablations.

Critical strengths
  1. The method is well motivated, using pixel-space modeling and orthogonal target decomposition to address the trade-off between structural fidelity and texture generation.
  2. OFD assigns complementary targets to DiT and PiT through a simple design. Ablations using the same architecture and parameter count support the effectiveness of this branch assignment.
  3. Experiments cover multiple super-resolution datasets and combine fidelity metrics, perceptual quality metrics, and visual comparisons, providing a multidimensional evaluation of restoration quality.
Critical weaknesses
  1. PCD overlaps with the uncited PixRestore [1] in pixel-space conditional flow modeling, input concatenation, and DINO guidance. The DiT/PiT architecture builds on PixelDiT [2], while the asymmetric target, orthogonal decomposition, and Procrustes-based projection are rooted in AsymFlow [3]. The novelty appears incremental, primarily residing in the OFD branch assignment rather than the underlying architecture or mathematical decomposition.
  2. The evaluation lacks several representative recent baselines. Beyond PixRestore [1], it omits the one-step methods ODTSR (CVPR 2026) [4] and StrSR (ECCV 2026) [5], as well as InvSR (CVPR 2025) [6], which supports both one-step and multi-step sampling.
  3. Orthogonality in output space does not imply conflict-free optimization. PiT depends on DiT features, so the two branches may still interact through shared parameters. The existing ablations support the effectiveness of the decomposition but do not establish gradient decoupling or confirm that the gains arise from the claimed semantic–texture specialization.
  4. The implementation details behind Figure 6 are ambiguous. If Eq. (18) is applied throughout sampling initialized from Gaussian noise, setting α=0\alpha=0 would preserve the initial noise in the orthogonal complement. This does not readily explain the smooth reconstructions shown in the figure and requires further clarification.

References:

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

[2] Yongsheng Yu et al. PixelDiT: Pixel Diffusion Transformers for Image Generation. CVPR, 2026.

[3] Hansheng Chen et al. Asymmetric Flow Models. arXiv preprint arXiv:2605.12964, 2026.

[4] Yushun Fang et al. One-Step Diffusion Transformer for Controllable Real-World Image Super-Resolution. CVPR, 2026.

[5] Jingkai Wang et al. Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution. ECCV, 2026.

[6] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps Image Super-resolution via Diffusion Inversion. CVPR, 2025.

Questions for authors
  1. Which designs are newly introduced in this work relative to PixRestore, PixelDiT, and AsymFlow? Could you clearly distinguish the contributions of existing mathematical formulations, architectural adaptations, and OFD branch assignment?
  2. Could you add comparisons with PixRestore, ODTSR, StrSR, and InvSR under a unified evaluation protocol, reporting sampling steps, runtime, and model size?
  3. Could you validate the necessity of the proposed branch assignment through target swapping, gradient isolation, or gradient conflict measurements, and clarify the scope of “decoupling” accordingly?
  4. At which sampling steps is α\alpha applied in Figure 6? When α=0\alpha=0, how is the orthogonal-complement component of the initial noise handled? Do the displayed images correspond to the final sampling state or a separately computed clean-image estimate?

If the authors adequately address my concerns, I will raise my rating.

10 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: Distinction from Related Works and Novelty

Answer: Thanks for your valuable comment. We would like to further clarify how our work differs from related works and highlight its core contributions.

(1) Distinction from the Concurrent Work PixRestore. PixelSR and PixRestore [1] both adopt pixel-space conditional flow modeling but differ in focus. Specifically, PixRestore supports unified image restoration through adaptive hierarchical visual guidance, whereas our work focuses on Real-world Image Super-Resolution (Real-SR). Through OFD, we align complementary prediction targets with the modeling characteristics of DiT and PiT to coordinate structure preservation and detail generation. We have added quantitative comparisons with PixRestore and related discussion.

(2) Utilization of Existing Foundations. We wish to clarify that we adopt only the PiT architecture from PixelDiT [2] and the asymmetric velocity target and orthogonal decomposition from AsymFlow [3]. Our key insight is to go beyond combining these components: we align the complementary targets produced by the orthogonal decomposition with the global and local modeling characteristics of DiT and PiT, respectively. DiT predicts semantic structure, while PiT models texture details. In this way, OFD turns architectural complementarity into task-aligned branch supervision, coordinating structure preservation with detail generation.

(3) Core Novelty of Our Work. Unlike latent-space methods that rely on VAE encoding and decoding, our method establishes a VAE-free, end-to-end, single-stage pixel-space framework for Real-SR that directly optimizes the final restored image. Through task-aligned branch supervision, OFD coordinates structure preservation and texture generation, offering a principled way to mitigate the longstanding trade-off between structural fidelity and perceptual realism.

References

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

[2] Yongsheng Yu et al. PixelDiT: Pixel Diffusion Transformers for Image Generation. CVPR, 2026.

[3] Hansheng Chen et al. Asymmetric Flow Models. arXiv preprint arXiv:2605.12964, 2026.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

Weakness 2 & Question 2: Comparisons with Representative Recent Baselines.

Answer: Thanks for your helpful suggestion. As suggested, we have added comparisons with PixRestore [1], ODTSR [4], StrSR-FLUX [5], and both one-step and five-step InvSR [6] under a unified evaluation protocol. Table R1 reports restoration quality, sampling steps, parameter counts, and runtime, with timing measured on identical hardware at the same input resolution. PixelSR achieves the best aggregate rank across the eight quality metrics on each benchmark. To ensure a fair comparison, we exclude PixRestore from the ScreenSR evaluation because it uses ScreenSR data for training, which may overlap with our test set. The updated complete table and related discussion have been added to the revised manuscript.

Table R1: Quantitative comparison on LSDIR, RealSR, ScreenSR, and UGCSR. Bold indicates the best result, and italics indicate the second-best result.

DatasetMethodStepsParameters (M)Inference Time (s) ↓PSNR ↑SSIM ↑LPIPS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑DeQA ↑Rank ↓
LSDIRODTSR11231.200.3420.190.52600.232348.984.4973.810.49733.942
StrSR1338.200.1719.610.52570.238751.503.6272.300.49823.993
InvSR-1133.840.0819.530.49940.268069.014.1773.620.52263.974
InvSR-5533.840.1818.740.48380.307675.514.5472.320.46083.676
PixRestore1726.100.1521.420.55580.274139.814.7263.380.32283.665
PixelSR (Ours)25773.404.4520.690.52700.220340.504.2573.190.56614.091
RealSRODTSR11231.200.3424.440.71590.2427102.275.9268.220.45363.563
StrSR1338.200.1722.280.68690.2729111.394.7868.590.47893.602
InvSR-1133.840.0823.850.69670.2891140.535.5768.570.45863.475
InvSR-5533.840.1822.770.66150.2966144.575.9965.880.41093.206
PixRestore1726.100.1527.410.76950.165733.746.1456.280.31663.004
PixelSR (Ours)25773.404.4825.340.71660.2922112.895.5769.590.53813.611
ScreenSRODTSR11231.200.3322.000.66060.197963.714.8172.690.51853.882
StrSR1338.200.1821.670.68900.202169.804.4572.570.47293.783
InvSR-1133.840.0821.300.60960.252289.814.1471.800.47183.814
InvSR-5533.840.1820.190.58760.297799.904.6469.430.40983.455
PixelSR (Ours)25773.404.4823.560.71560.208269.134.3372.860.48243.831
UGCSRODTSR11231.200.3322.440.69830.194775.725.4074.050.51213.914
StrSR1338.200.1823.660.78350.168765.805.5374.730.55494.062
InvSR-1133.840.0822.070.66450.242999.195.1373.860.49613.755
InvSR-5533.840.1821.290.64710.2809106.035.5771.860.44993.496
PixRestore1726.100.1525.980.80610.124563.214.9070.630.45513.663
PixelSR (Ours)25773.404.4425.720.80560.145956.224.8475.690.56123.951

References

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

[2] Yongsheng Yu et al. PixelDiT: Pixel Diffusion Transformers for Image Generation. CVPR, 2026.

[3] Hansheng Chen et al. Asymmetric Flow Models. arXiv preprint arXiv:2605.12964, 2026.

[4] Yushun Fang et al. One-Step Diffusion Transformer for Controllable Real-World Image Super-Resolution. CVPR, 2026.

[5] Jingkai Wang et al. Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution. ECCV, 2026.

[6] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps Image Super-resolution via Diffusion Inversion. CVPR, 2025.

AuthorsOpenReview ↗

Official Response by Authors – Part 3

Weakness 3 & Question 3: Validation of Task-Aligned Branch Assignment and Clarification of Decoupling

Answer: Thank you for this constructive feedback. We would like to clarify that Orthogonal Flow Decoupling (OFD) is designed to align complementary semantic and texture prediction targets with the modeling characteristics of DiT and PiT, respectively. Here, “decoupling” refers to the separation of prediction targets, rather than complete gradient independence on shared parameters.

Following your suggestion, we have added a target swapping ablation under the same architecture and training settings and separately compared gradient cosine similarities and conflict rates on shared parameters across OFD, target swapping, and the undecomposed baseline.

As shown in Table R2, OFD achieves better restoration quality and a lower negative-cosine rate in the gradient measurements, supporting the effectiveness of the proposed branch assignment in terms of both performance and optimization compatibility. We have incorporated these results and analyses into the revised manuscript and adjusted the relevant statements accordingly.

Table R2: Effectiveness Analysis of Task-Aligned Branch Assignment. Bold indicates the best result.

MethodPSNR ↑SSIM ↑LPIPS ↓NIQE ↓MUSIQ ↑MANIQA ↑DeQA ↑FID ↓Mean Gradient CosineNegative Cosine Rate (%) ↓Mean Negative Magnitude (×10⁻³) ↓
uAu_{\mathrm A}24.790.68450.31575.3869.570.46803.55117.920.22015.757.60
Target Swapping24.880.69270.31614.7169.410.51803.49118.150.169210.578.50
OFD (Ours)25.340.71660.29225.5769.590.53813.61112.890.24183.853.30

Weakness 4 & Question 4: Implementation Details of α\alpha in Figure 6

Answer: Thank you for your valuable feedback. Figure 6 examines the contribution of the orthogonal-complement branch to texture and frequency responses at the final sampling step, with the preceding restoration state held fixed. Specifically, the first 24 steps use the full prediction (α=1\alpha=1), and α\alpha is varied only at step 25. Thus, setting α=0\alpha=0 disables only the final orthogonal-complement update, retaining the component already denoised over the preceding steps rather than the initial Gaussian noise. This explains the smooth reconstructions shown in Figure 6. The displayed images correspond to the final sampling states, not separately computed clean-image estimates. We have clarified these details in the revised manuscript.

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 PZWQ acknowledged that the additional evidence and clarifications strengthened the paper, Reviewer Evyq confirmed that the new experiments addressed some of their concerns, and Reviewer pdV1 recognized improvements in both the supporting evidence and the clarity of our contributions. Below, we provide a brief summary of the author–reviewer discussion.

  • Reviewer PZWQ recognized that our method is “well motivated” and that controlled ablations support the effectiveness of OFD’s branch assignment. We added recent baselines, target-swapping experiments, and gradient-conflict analysis, which the reviewer acknowledged “have clearly strengthened the paper.” Following our clarification of Eq. (18), the reviewer also confirmed that the Figure 6 concern was resolved and appreciated our “careful checking” and engagement throughout the discussion. The corresponding corrections have been incorporated into the updated manuscript.

  • Reviewer Evyq recognized the well-motivated problem, the unified pixel-space restoration framework, and the clear rationale for separating global generation from local reconstruction. We supplemented target-swapping and denominator-threshold experiments, expanded fidelity–realism comparisons, and clarified UGCSR construction and training–test independence. The reviewer subsequently confirmed that “The authors have addressed some of my concerns with additional experiments” and stated that this evidence would be taken into account.

  • Reviewer pdV1 praised the “particularly well-controlled evaluation of Orthogonal Flow Decoupling”, noting that matched architecture, parameter count, and training settings make the observed gains convincing. We added gradient measurements, VAE reconstruction analysis, and component comparisons to clarify the contribution of task-aligned supervision. The reviewer acknowledged that these analyses “have helped clarify the contributions of the proposed method” and concluded that “the new experiments have strengthened the paper.”

Compared with latent-space approaches, our PixelSR avoids VAE reconstruction errors through direct pixel-space optimization. OFD further aligns the physical properties of complementary prediction targets with the global and local modeling capabilities of DiT and PiT, coordinating structural recovery and texture generation.

We hope that the additional evidence, clarifications, and manuscript improvements will support your consideration of our work. Once again, we sincerely thank the reviewers and area chairs for their valuable time and constructive feedback.

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.