- 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
-
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.
-
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.
-
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.
-
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.
-
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.
-
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
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.
| Method | Task | Prediction target | Global/local modeling | Low-rank projection | LR + DINOv2 guidance | OFD | Parameters | Training data |
|---|---|---|---|---|---|---|---|---|
| PixelDiT-XL/16 | Image generation | ✓ | ✗ | ✗ | ✗ | 797M | ImageNet | |
| DiP-XL/16 | Image generation | ✓ | ✗ | ✗ | ✗ | 631M | ImageNet | |
| AsymFlow-H/16 | Image generation | ✗ | ✓ | ✗ | ✗ | 953M | ImageNet | |
| PixelSR (Ours) | Real-SR | to DiT; to PiT | ✓ | ✓ | ✓ | ✓ | 773.4M | LSDIR, 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 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.
| Method | NFE | PSNR ↑ | MUSIQ ↑ | Time (s) ↓ | Params (M) ↓ |
|---|---|---|---|---|---|
| StableSR | 200 | 24.40 | 70.82 | 5.71 | 1540 |
| DiffBIR | 50 | 23.08 | 74.01 | 5.19 | 380† |
| DiT4SR | 40 | 21.74 | 72.94 | 5.17 | 21380 |
| VOSR | 25 | 21.89 | 74.74 | 3.86 | 619.4 |
| DiT-SR | 15 | 23.96 | 69.93 | 1.87 | 61 |
| OSEDiff | 1 | 22.40 | 73.58 | 0.46 | 1760 |
| SinSR | 1 | 23.82 | 69.27 | 0.33 | 174 |
| CODSR | 1 | 22.23 | 74.98 | 6.21 | 35.79 |
| PixelSR (Ours) | 25 | 25.72 | 75.69 | 4.44 | 773.4 |
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 and generate 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 super-resolution using spatially aligned LR–HR image pairs, with resolutions of and , 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 super-resolution setting, whereas ScreenSR and UGCSR use 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 on RealSR by varying only at the final sampling step, while keeping for the preceding 24 steps. Table R5 reports PSNR, SSIM, LPIPS, MUSIQ, and MANIQA. Increasing 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 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 on RealSR. Bold indicates the best result.
| PSNR ↑ | SSIM ↑ | LPIPS ↓ | MUSIQ ↑ | MANIQA ↑ | |
|---|---|---|---|---|---|
| 0 | 21.93 | 0.6143 | 0.3345 | 59.44 | 0.3715 |
| 0.5 | 24.89 | 0.7014 | 0.3083 | 65.84 | 0.4533 |
| 1 | 25.34 | 0.7166 | 0.2922 | 69.59 | 0.5381 |
| 1.5 | 23.26 | 0.6327 | 0.3318 | 72.67 | 0.5567 |
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.
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 uA 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.
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 v-prediction unchanged, before progressively introducing x0-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.