acceptodds
Withdrawn from ICLR 2027

WANSR: Distilling Video Diffusion for Trajectory-Aware Real-World Image Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world image super-resolution (Real-ISR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs perturbed by unknown and mixed degradations. Although diffusion-based methods enable generating perceptually satisfying results, their intermediate denoising states are typically treated as numerical variables for sampling, making it hard to interpret and explicitly constrain the restoration process. Therefore, we ask the question: Can the intermediate restoration process be explicitly modeled? In this paper, we propose WANSR, a trajectory-aware framework that reformulates Real-ISR as a conditional video generation problem. Instead of directly predicting a single HR image, WANSR models a sequence of restoration states that explicitly describes the progressive evolution toward the final high-quality distribution. To stabilize this evolution, we introduce anchor-guided trajectory alignment by aligning transition directions and anchor-level magnitudes. We further propose quality-verified trajectory refinement, which evaluates restoration progress of each latent state and enforces monotonic quality improvement along the generated trajectory. Extensive experiments on multiple Real-ISR benchmarks demonstrate that our WANSR achieves superior reconstruction performance, while explicitly revealing the progressive process of degradation removal and details reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Real-world image super-resolution (Real-ISR), Video Generation, Diffusion

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_fxfMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

For image SR using diffusion, authors propose to generate Y (SR) from X (LR) by controlling the trajectory of successive improvement rather that taking any arbitrary trajectory to reach Y from X. They also show that the no of steps required for diffusion is less compared to competing method.

Critical strengths
  1. Less number of steps for diffusion
  2. Good number of experimentation
  3. Better SOTA for reference based image quality metrics
Critical weaknesses
  1. It is not at all clear why should the evolution in restoration/deblurring be along a specified trajectory? The trajectory gets decided by the anchors. It also appears that the guided path is more or less linear locally -- suggesting the manifold of the latent space is Euclidean. In general the manifold may follow quite a bit of geodesic evolution which your method does not take care. That is why you are possibly seeing a reasonable improvement in the reference-based metrics like PSNR where the Euclidean distance is finally what is matched, while it is performing poorly for no-reference quality metrics.

  2. Why should I even care that X -> Y evolution must pass through a path Z? We want good results and it is the end result that you are comparing for all baseline and other methods. We do not even know if the path Z is good.

  3. Despite distillation, your parameter space is very huge compared to competitors and so is the FLOP count. Then how come your latency is much lower. Is there something we are missing. What is your definition of latency.

  4. It is not clear how is the fake distribution p_{fake} estimated.

Questions for authors

see above.

Reviewer_Rjf8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WANSR leverages a video diffusion model to generate a sequence of restoration states for single-image restoration, explicitly modeling the intermediate recovery process. A pretrained state judge is introduced to guide the intermediate latent states to evolve monotonically toward the high-resolution (HR) distribution.

Critical strengths

The method creatively exploits the intermediate states in the Real-ESRGAN degradation pipeline as restoration anchors. By reversing these states, it reformulates single-image restoration as an explicit sequence of progressive recovery steps, making the intermediate restoration process observable.

Critical weaknesses

The computational cost is not sufficiently justified by the reported quality improvements. WANSR employs 5.704B parameters and jointly processes 11 temporal latent states to restore a single image, making it substantially more expensive than the compared one-step and few-step models. Comparisons with stronger multi-step restoration methods, such as DreamClear and LucidFlux, are needed to better assess the benefits of such a large video backbone. Moreover, Table 1 does not demonstrate a favorable fidelity–perception trade-off. While WANSR achieves higher fidelity and better reference-based similarity, its no-reference perceptual scores are considerably worse. On RealSR, its CLIP-IQA is only 0.5290, compared to 0.7166 for TSD-SR. Meanwhile, Table 2 reports a latency of 42 ms for TSD-SR versus 620 ms for WANSR, making TSD-SR approximately 15× faster. These results raise concerns about whether the fidelity gains justify the degraded no-reference perceptual quality and substantially higher computational cost.

Questions for authors

Does WANSR use the same 41-frame sequence for all inputs, regardless of degradation type or severity? Given that the training anchors are constructed by reversing successive degradation operations in the Real-ESRGAN pipeline, how does WANSR adapt its restoration trajectory to inputs with only a single degradation type or mild degradation? Please provide results on such cases to demonstrate whether the model can restore them without introducing unnecessary changes.

Reviewer_J9AcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WANSR, a trajectory-aware framework for real-world image super-resolution (Real-ISR) that reformulates restoration as conditional video generation. It constructs pseudo-restoration trajectories by reversing synthetic degradation sequences and introduces three components: progressive latent evolution, anchor-guided trajectory alignment (direction and magnitude alignment), and quality-verified trajectory refinement using a state judge. The method aims to improve restoration fidelity and controllability while producing intermediate restoration states. Experiments show competitive reconstruction performance on synthetic and real-world benchmarks, although trade-offs between distortion and perceptual quality remain.

Critical strengths

1.Reformulating Real-ISR as progressive latent evolution provides interpretable intermediate states and greater control over the restoration process than endpoint-only prediction. 2.The method achieves strong reconstruction fidelity on the evaluated benchmarks.

Critical weaknesses

1.The trajectories are constructed using the synthetic Real-ESRGAN degradation pipeline and pixel-wise interpolation. These intermediate states may not accurately represent real-world restoration dynamics, especially under synthetic-to-real degradation mismatch. Since all three trajectory-related components rely on these pseudo-trajectories, their biases may propagate throughout training. 2.The state judge is trained and evaluated against pseudo-trajectory ordering. Monotonically increasing scores therefore do not necessarily imply monotonically improving perceptual quality or structural fidelity on real images. Independent validation of the judge is lacking. 3.Improvements in final-image metrics do not establish that intermediate states follow meaningful restoration dynamics. The evaluation also provides limited evidence of robustness to diverse real-world degradation distributions. 4.Despite using only four denoising steps, WANSR has 5.7B parameters and substantial computational requirements compared with lightweight one-step alternatives, which may limit practical deployment.

Questions for authors

1.How do the authors justify the use of reversed synthetic degradation sequences and pixel-wise interpolation as meaningful restoration supervision? Could multiple degradation trajectories or uncertainty-aware weighting improve robustness? 2.How does WANSR perform when the test-time degradation distribution differs substantially from the Real-ESRGAN training pipeline? Can the authors provide controlled mismatch experiments?

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

Related papers

Discussion (0)

Sign in to comment.