acceptodds
Withdrawn from ICLR 2027

Rethinking Fidelity–Perception Trade-off in Latent-Free Super-Resolution via Pixel Flow

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution observations. Recent models using diffusion models and flow matching have shown strong perceptual restoration ability, yet we observe two key limitations that hinder their robustness and fidelity: 1) their performance is sensitive to different noise and sampling stochasticity; 2) per-step noise prediction may fail under composite degradations and accumulate residual errors, causing identity, texture, and structural drift. To address these issues, we rethink the fidelity–perception trade-off from a latent-free perspective and propose PixelSR, a pixel-flow SR framework that directly models the restoration trajectory in image space. Instead of predicting per-step noise in a compressed latent space, PixelSR predicts clean high-resolution images and derives the corresponding pixel-space flow for iterative refinement, thereby anchoring the generation process to the high-quality image manifold while preserving low-level structures from the input. This clean-data-centric formulation reduces error accumulation and provides a more stable path toward perceptually realistic yet faithful reconstruction. Interestingly, our experiments reveal that varying the number of sampling steps provides a simple and effective mechanism to navigate the fidelity–perception trade-off. Extensive experiments on synthetic and real-world SR benchmarks demonstrate that PixelSR achieves a better fidelity–perception balance than strong diffusion-based baselines, with improved identity preservation, structural consistency, and perceptual quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
low-level vision, flow matching, super resolution, image restoration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pVqEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PixelSR, a real-world image super-resolution framework that models restoration as a deterministic pixel-space flow from a low-quality image to a clean high-resolution image. The model predicts the clean image and derives the flow for iterative refinement. The authors also propose using the number of sampling steps to control the fidelity–perception trade-off. Experiments evaluate PixelSR on synthetic and real-world benchmarks and compare it with several diffusion-based restoration methods.

Critical strengths
  1. The paper explores a clear alternative to noise-centric restoration by directly predicting clean images in pixel space. Its deterministic trajectory is a potentially useful way to reduce sensitivity to random noise and preserve input structure.

  2. The evaluation covers synthetic and real-world benchmarks, includes full-reference and no-reference metrics, and studies how the number of sampling steps affects the results.

Critical weaknesses
  1. The distinction from related flow-based SR methods is not sufficiently established. The paper discusses FlowSR (ICCV 2025) and FluxSR (ICML 2025), which also use flow-based trajectories for super-resolution, but does not include them in the main comparison. A clearer technical comparison is needed to support the novelty and performance claims.

  2. Algorithm 1 contains an apparent typographical error. Its interpolation weights are reversed relative to Equation (1) and the inference procedure.

  3. The sampling-step experiments do not consistently support a progressive fidelity–perception trade-off. For example, on DRealSR, the 10-step result outperforms the 50-step result on all reported metrics, making the effect of sampling-step count on this trade-off unclear.

  4. The contribution of direct clean-image prediction is not isolated. The ablations do not compare direct HQ-image prediction with velocity prediction using the same backbone and trajectory, so the gains attributable to this prediction formulation remain unclear.

  5. The claim that the trajectory remains on the image manifold is insufficiently supported. Linear interpolation between LQ and HQ images does not by itself establish that intermediate states remain on the natural-image manifold. This claim requires further explanation or supporting evidence.

Questions for authors
  1. How does PixelSR differ technically from FlowSR and FluxSR?

  2. Could you compare clean-image prediction with velocity prediction using the same backbone and trajectory? This would help assess the contribution of the proposed prediction formulation.

  3. On DRealSR, the 10-step setting outperforms the 50-step setting across all reported metrics. How do you interpret this result, and how does it support the claimed fidelity–perception trade-off?

  4. What do you mean by the claim that the interpolation trajectory remains on the image manifold?

Reviewer_a8SdOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PixelSR, a latent-free pixel-flow framework for real-world image SR. The authors argue that existing diffusion- and flow-based methods can be sensitive to sampling randomness and can accumulate prediction errors under complex degradations. PixelSR combines a flow-matching loss with LPIPS and DINO feature losses. The authors also claim that changing the number of sampling steps provides a way to control the fidelity–perception trade-off.

Critical strengths
  1. Directly predicting a clean image rather than making noise or velocity the primary prediction target is easy to understand. But it was proposed by JIT, which is not very novel.

  2. The paper reports both full-reference and no-reference metrics and studies different sampling-step settings. This provides initial evidence about the fidelity–perception trade-off.

Critical weaknesses
  1. Equation (1) defines t=0 as the LQ image and t=1 as the HQ image. However, Algorithm 1 reverses this direction. Equation (8) also conflicts with the earlier definition.

  2. PixelSR requires 50 diffusion steps, which is heavy. Inference time and model parameters comparisons should be given.

  3. Nolvelty is my main concern. In addition, the performance improvement is not clear.

Questions for authors

The paper claims a better fidelity–perception balance, but PixelSR does not lead on several metrics in the tables. How is “better balance” defined? Can the authors narrow this claim and specify the metrics and criteria used for the comparison?

Reviewer_WgHXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper trains a latent-free model of FlowSR [1] with pixel-level losses as LPIPS and DINO loss. The performance of their method is comparable to existing diffusion or flow-based super-resolution (SR) methods. Furthermore, they observed that the number of flow-steps that is used at the inference acts as a knob that can control whether to maximize pixel fidelity or realism (perception).

References:

[1] Xu, Jiaqi, et al. "Fast image super-resolution via consistency rectified flow." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

Critical strengths
  1. They propose a concise pipeline where no latents are involved and no pretrained large models are used. Due to the deterministic nature of their method, the restored images have stable reconstruction performances compared to noise-involved diffusion/flow models.

  2. They make an interesting finding where the number of flow-steps that is used at the inference acts as a knob that can control whether to maximize pixel fidelity or realism (perception), which is not shown in FlowSR [1] paper.

References:

[1] Xu, Jiaqi, et al. "Fast image super-resolution via consistency rectified flow." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

Critical weaknesses
  1. Some parts of this paper's writing is highly misleading. First, the idea of the Eq (1) is from FlowSR [1] and thus should be stated so. This results in overstating the novelty of the formulation and the method that this paper proposes. Furthermore, [1] is not reported in any of their benchmark tables. Note that the whole picture of thinking the path between x_L to x_H relates to more of prior works [2, 3] that are not referenced in the paper. Furthermore, one of the most important paper about the perception-distortion tradeoff of inverse problems [4] is not cited at all.

  2. The rationale behind the tradeoff is not found on the paper. According to Table 2., the tradeoff does seem to appear, but it is only an observation. The readers are left confused about why more steps gives better no-reference image quality (NR-IQA) scores and give worse full-reference image quality (FR-IQA) scores. Furthermore, it is hard to understand why more steps can hurt both FR/NR-IQA scores. It is noteworthy that the perception-distortion tradeoff of image restoration is not something that is only empirical - it is rather a hard theoretical bound that is derived.

  3. One of the reasons why latent diffusion models are often favored over pixel-level diffusion models is due to their compute efficiency. Therefore to justify the latent-free design pipeline, the compute efficiency of the proposed pipeline needs to be reported. However, this is not reported. In Table 1., the 50 step version of their method is reported, which is compared to various 1 step version of other methods. Furthermore, FlowSR [1] is not reported on that table, but beats this paper's method on various metrics according to the reports in [1]. This highly question their motivation to choose a latent-free pipeline design.

  4. The authors considers the removal of stochasticity only as a good thing and doesn't mention what happens if the stochasticity of fully removed from a pipeline.

  5. Qualitative experimental results are too limited and even the only existing one (Figure 5.) doesn't show much differences. There are no qualitative results on the tradeoff that they claim. Furthermore, the ablative study is too limited. Testing on the latent version of their method, ablation on training loss hyperparameters, etc will make the foggy picture of this paper clearer.

References:

[1] Xu, Jiaqi, et al. "Fast image super-resolution via consistency rectified flow." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[2] Park, Seobin, and Tae Hyun Kim. "Progressive image super-resolution via neural differential equation." ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022.

[3] Delbracio, Mauricio, and Peyman Milanfar. "Inversion by direct iteration: An alternative to denoising diffusion for image restoration." arXiv preprint arXiv:2303.11435 (2023).

[4] Blau, Yochai, and Tomer Michaeli. "The perception-distortion tradeoff." 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2018.

Questions for authors

See weakness. Especially, please give detailed explanation about 2.

I am open to raise my score if all my concerns are resolved.

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

Related papers

Discussion (0)

Sign in to comment.