acceptodds
Withdrawn from ICLR 2027

What Does High-Resolution Supervision Actually Teach? Separating Source Support from Resolution Specialization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-resolution super-resolution corpora are motivated by the assumption that natively high-resolution images supply supervision that lower-resolution data cannot. Existing comparisons cannot test it: replacing a low-resolution corpus with a native-4K one changes the degradation operator, the source pixels available for sampling, the number of scenes, the field of view of a fixed crop, the allocation of compute and the train–test match all at once, so the reported difference is a sum of six effects and identifies none of them. We measure the raster’s own contribution using same-source resolution pyramids, one declared degradation, a fixed optimisation budget, and evaluation of every arm both on a common test tier and across tiers. Source support, scene count and per-scene coverage are bound by one constraint, so no design varies the raster alone; we run all three that release one of them. On a source-rich corpus the apparent 4K advantage — 0.835 dB at 43M parameters — survives none of them, at any capacity or scale factor. On two unrelated RAW-derived corpora the preference between training rasters instead reverses with the target resolution. High-resolution data therefore supplies either additional source support or closer alignment with the deployment resolution, and the nominal raster measures neither.

Primary area:
datasets and benchmarks
Keywords:
image super-resolution, training data, source support, resolution specialization, controlled evaluation

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_z4aqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether native high-resolution supervision provides a distinct benefit for single-image super-resolution, or whether reported gains from high-resolution datasets are mainly due to confounded factors. The authors argue that replacing a low-resolution corpus with a native-4K corpus changes many variables at once, including nominal raster, source-pixel support, scene count, crop field of view, degradation operator, optimization budget, and train-test resolution match.

The paper proposes a controlled evaluation protocol based on same-source resolution pyramids, common-tier evaluation, source-support matching, iso-scene controls, and cross-tier interaction measurements. The main empirical result is that on DIV8K, an apparent 4K advantage disappears once source-pixel support is matched. However, on RAW-derived corpora such as RAISE and FiveK, the preferred training raster changes with the target resolution, suggesting that high-resolution supervision can still matter through corpus-specific train-test resolution alignment.

Critical strengths

The paper raises a relevant issue in super-resolution evaluation: nominal dataset resolution is confounded with source support, scene count, crop field of view, degradation choices, and train-test resolution alignment. This is an important point, and the paper is correct that conventional dataset swaps do not isolate the effect of high-resolution supervision.

The authors make a serious attempt to control these factors using same-source pyramids, common-tier evaluation, source-support matching, and cross-tier measurements. The operator audit and appendix controls also show care in experimental design.

Critical weaknesses
  1. The paper is very difficult to read, which substantially weakens its contribution. The writing is dense, highly formal, and often prioritizes abstract notation and identification language over a clear explanation of the experimental logic. Even after careful reading, it is not always easy to distinguish the main claim, the exact estimand being tested, and the practical implication of each control experiment. Since the paper’s contribution is primarily an analysis/evaluation study rather than a new model, clarity is especially important. In its current form, the presentation makes the core message unnecessarily hard to extract.

  2. The central causal interpretation is not convincing enough to support the paper’s claims. The paper argues that the apparent 4K advantage disappears once source support is matched, but the matching procedure necessarily changes other important factors, especially scene count and/or per-scene coverage. The authors acknowledge that nominal raster cannot be varied while holding source support, scene count, and scene coverage all fixed, but this is precisely the problem: the proposed experiments do not cleanly isolate a pure high-resolution supervision effect. As a result, the main conclusion rests on controlled comparisons that still have unavoidable and consequential confounds.

  3. The paper stops mostly at analysis and does not provide enough actionable insight for the SR community. Given that training with 4K images can introduce substantial storage, sampling, and computation costs, the observation that high-resolution supervision may be inefficient or only conditionally useful is important. However, the paper does not go further to propose reliable training strategies, data selection rules, adaptive sampling schemes, or test-time adaptation methods that could help practitioners use this observation. This limits the novelty and impact of the work: the paper identifies a systematic issue, but does not sufficiently show how to address it.

I used an AI assistant to help organize and polish parts of my review after reading the paper. The assistant helped me formulate one concerns that I agree with critical weakness (2).

I reviewed and edited the final review myself.

Besides, I recommend that the authors cite Improving Image Restoration by Revisiting Global Information Aggregation (ECCV 2022).

Questions for authors

If natural predictors fail to determine when high-resolution supervision matters, what concrete guidance does the paper provide to practitioners beyond running the proposed measurement for every new corpus?

Can the authors propose or evaluate any practical strategy, such as data selection, adaptive raster sampling, curriculum training, or test-time adaptation, that would help reduce the cost of 4K supervision while preserving its benefits when it is actually useful?

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

The paper investigates whether natively high resolution data provides unique supervision training signal compared to lower resolution. There are six hidden confounders in some datasets comparisons, making it difficult to determine the source of the gains derived from training on high resolution datasets. To isolate the causal effect, the paper presents a rigorous identification protocol and experimental demonstration with some key findings:

In the DIV8K evaluation, the apparent 4K advantage in uncontrolled setups can be fully explained by source-pixel support. Once source pixels are matched, however, this advantage disappears across all scale factors. When a fixed pixel budget is allocated to more scenes at lower resolution, it consistently outperforms allocating the budget to fewer scenes at 4K. However, on RAW-derived corpora, the preference between training resolution demonstrates that resolution specialisation can occur depending on the source. Finally, the paper attempts to predict whether a dataset exhibits resolution specialisation.

Critical strengths
  1. The theoretical analysis is sound. By formalizing the fundamental data budget constraint as S = N·a·|R|, the entire theoretical analysis is coherently bound and the three single-release configurations (unmatched, support-matched, and iso-scene) are naturally derived. Additionally, appropriate mathematical propositions are introduced to provide a theoretical basis for the entire protocol's construction.

  2. The experimental investigation is meticulous. The paper exposes an overlooked confound in the existing literature by carefully reviewing the implementation code. Based on these findings, executes a strictly controlled protocol involving same-source pyramids, committed operators, explicit separation of seed versus subset variability, and cross-tier interaction evaluation. By debunking the universal 4K prior myth and uncovering resolution specialization in RAW domains, this work makes an advance for the super resolution community.

Critical weaknesses
  1. The paper introduces several notations and terms whose definitions may differ from those in the normal image restoration usage. This creates cognitive friction while reading: The symbol I is universally used as shorthand for an image tensor; Exposure usually refers to photometric sensor exposure; Similarly, standard image resolution and dataset size are redefined in terms of nominal raster R and source support S. Perhaps including a notation table that explicitly defines each terms and notations would improve the reading experience.

  2. The core conclusions are based on regression objectives. Although the authors include a preliminary GAN experiment using ESRGAN in Appendix F, the SISR has shifted toward generative restoration paradigms powered by large scale diffusion priors (e.g., DiffBIR, HYPIP). In generative super resolution, networks operate on vastly larger parameter regimes and reward unpredictable, high-frequency hallucinations. It remains uncertain whether the source-support explains merely an artifact of the regression paradigm.

Questions for authors

The same-source pyramid is central to your identification process. While the 4K tier is a faithful capture, the 1K and 2K tiers are generated via synthetic bicubic decimation with antialiasing. In a real-world scenario, a native 1K image is subject to distinct camera sensors and image signal processor pipelines that differ fundamentally from idealized mathematical downsampling. Could the authors clarify why lower tiers that are digitally downsampled can fairly proxy "native lower-resolution photographs"? Could the disappearance of the 4K tier on DIV8K be partly induced by the mathematical self-similarity enforced by synthetic downsampling?

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

This paper explored the contribution of high resolution in the training dataset in the super-resolution task. It challenges the common assumption in super-resolution research that high-resolution training data naturally provides better supervision. The paper constructs the same-source resolution pyramid, fixes the degradation operator and optimization budget, and discusses the constraints of source support, scene count and per-scene coverage respectively.

The paper shows that the surface advantage of native-4K data vanishes after matching the pixel budget for source-rich scenes, and that more low-resolution scenes with a fixed budget may outperform higher raster. The paper argues that the benefit of high-resolution data is in providing additional data sources or dimensions that are more aligned with the test dataset, rather than nominal raster itself.

Critical strengths
  1. The problem is clearly defined and has community value. It explores the contribution of the nominal raster in super-resolution datasets. The paper attempts to point out that the traditional comparison of replacing low-resolution datasets with native-4K datasets does not identify the separate contribution of resolution because a dataset swap changes multiple variables simultaneously. This criticism has practical relevance, especially for the long-standing experimental habit of relying on the dataset level comparison in the super-resolution field.

  2. The experiment is controlled systematically. The paper constructs homology resolution pyramid, fixes degradation operators, optimisation budget, and training settings to discuss the impact of high resolution, which is more convincing than the replacement of datasets.

Critical weaknesses
  1. The external validity of the conclusions is still limited. The paper achieves strong internal control through same-source resolution pyramids, a declared degradation operator, and fixed optimization budgets, but this setting may not fully represent real native-4K data acquisition, real camera pipelines, real-world degradations, or deployment distributions, suggesting that broader validation on more realistic native high-resolution corpora would be valuable.

  2. In addition, the paper includes capacity scaling and a shifted-window transformer perturbation, but it does not fully explore whether the conclusions remain stable across a wider range of SR models, such as Real-ESRGAN, HAT, or perceptual/GAN-based SR systems.

  3. The paper’s practical guidance needs to be further strengthened. While the conclusion that high-resolution supervision reflects either additional source support or better alignment with the deployment resolution is conceptually useful, the paper does not provide a reliable way to predict which regime a new corpus will fall into.

Questions for authors
  1. The conclusion of this paper is based on the experiment of same-source resolution pyramids, fixed degradation operator, but the real native-4K or high resolution images usually do not come from downsampling, can try to explain: Does the disappearance of the high-resolution advantage apply only to synthetic degradation, and can it be extrapolated to real native high-resolution datasets?

  2. The paper is based on the verification under EDSR and SwinIR. For the diffusion model, as well as the model with a larger number of parameters, whether the conclusion is consistent with the paper.

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

Related papers

Discussion (0)

Sign in to comment.