Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Learning Arbitrary-Scale RAW Image Downscaling with Wavelet-based Recurrent Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A wavelet-based recurrent reconstruction framework performs arbitrary-scale RAW image downscaling, including non-integer factors like 1.3x, with state-of-the-art fidelity.

desk verdict First learning-based arbitrary-scale RAW downscaling with a sensible wavelet design; the 1.3x SOTA claim is conditional on unvalidated ground-truth alignment. read the letter →

arxiv 2507.23219 v1 pith:7ZKBCOQ2 submitted 2025-07-31 eess.IV cs.CV

classification eess.IVcs.CV
keywords RAWimagedownscalingarbitrary-scalewavelettransformrecurrentreconstructionhigh-frequencypredictionReal-NIRDdatasetenergy-maximizationlossrescaling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to establish that RAW-domain image downscaling can be done at arbitrary scale factors, including non-integer ones like 1.3x, without the blur, aliasing, and ghosting that interpolation and sRGB-domain learning methods produce. It argues that by decomposing the high-resolution RAW image into wavelet sub-bands and recursively reconstructing a low-resolution version level by level, the method preserves both global structure and high-frequency texture. If correct, this would give cameras and imaging pipelines a practical way to shrink RAW data directly for storage or transmission, keeping more useful detail than current pipelines that must convert to sRGB first. The paper also contributes the Real-NIRD dataset of real 1.3x RAW pairs to benchmark non-integer downscaling.

What carries the argument

The central mechanism is the wavelet-based recurrent reconstruction decoder (WRRD), which uses the 2D discrete wavelet transform (with Haar wavelets) to split the HR feature map at each level into one low-frequency coefficient $A^k_{HR}$ and three high-frequency components $\{V^k_{HR}, H^k_{HR}, D^k_{HR}\}$. The LASDM downscales the low-frequency coefficient by expressing any rational scale factor as a fraction $n/m$ and applying pixel unshuffle and pixel shuffle operations to merge information in the channel dimension, handling both integer and non-integer downscaling without interpolation artifacts. The HFPM predicts the matching LR high-frequency sub-bands by keeping the top $n h'/m$ and $n w'/m$ sparse coefficients along vertical and horizontal directions and using a normalized cross-attention location map to remap them to the target resolution, and an energy-maximization loss $L_{em}$ pushes the energy of the predicted high-frequency maps toward that of the HR ones. These components are combined with hierarchical wavelet-consistency and content-consistency losses to train the network end to end.

What would settle it

A direct way to test the claim is to register the Real-NIRD pairs with a robust alignment step (for instance, RANSAC or optical flow on the full-resolution frames) and recompute PSNR, SSIM, and LPIPS for all methods on the aligned pairs; if the proposed method's margin over interpolation shrinks to near zero, part of the reported gains would be attributable to misalignment rather than downscaling quality. An equally decisive test is to synthesize HR-LR pairs by applying a known 1.3x downscale to a high-resolution RAW image and checking whether the method's gains persist when the ground-truth alignment is exact.

Watch

Extended reading notes

Core claim

The central claim is that a wavelet-based recurrent reconstruction decoder, combined with a low-frequency arbitrary-scale downscaling module (LASDM), a high-frequency prediction module (HFPM), and an energy-maximization loss, can reconstruct a low-resolution RAW image from a high-resolution one at any rational scale while retaining structural and textural fidelity. The method treats the target LR image as the low-frequency wavelet coefficient of the HR image and regenerates the missing high-frequency sub-bands at each level of a coarse-to-fine inverse wavelet transform. On the proposed Real-NIRD 1.3x dataset and on the Real-RawVSR 2x/3x/4x splits, the authors report the best PSNR, SSIM, and LPIPS among all compared methods, including interpolation baselines and six learning-based sRGB downscaling approaches. This makes it, by the paper's account, the first learning-based arbitrary-scale RAW downscaling framework to cover both integer and non-integer factors.

Load-bearing premise

The load-bearing assumption is that the 1.3x Real-NIRD image pairs are physically aligned, so the low-resolution photo is a faithful downscale of the high-resolution scene; if residual parallax or vibration shifts edges between the two shots, the reported quality gains may partly reflect alignment error rather than better downscaling.

Editorial extensions

If this is right

  • If the reported results hold, learned RAW-domain downscaling can outperform both traditional interpolation in RAW space and learning-based rescaling in sRGB space, giving better low-resolution images for storage and preview.
  • The framework extends cleanly to non-integer factors such as 1.3x, which interpolation-based pipelines handle poorly, potentially simplifying real-world capture workflows that need arbitrary output sizes.
  • The energy-maximization loss provides a simple, general training signal for preserving high-frequency content when information must be discarded.
  • The Real-NIRD dataset offers a public benchmark for non-integer RAW downscaling, enabling future methods to be compared on realistic 1.3x pairs rather than synthetic crops.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Since the Real-NIRD pairs were captured with a tripod and remote trigger but without explicit image alignment, part of the reported gains could come from residual parallax or vibration; a natural test is to register the pairs with a robust alignment step and re-measure all methods.
  • The same wavelet recurrent design could in principle be applied to video RAW downscaling, where temporal consistency across frames is an additional constraint the current framework does not address.
  • The energy-maximization loss might transfer to other image-to-image tasks that sacrifice resolution, such as deep learning-based downsampling for learned compression, where preserving salient high-frequency energy matters.
  • A direct synthetic evaluation with known ground-truth LR images would isolate downscaling quality from real-capture alignment issues, which the current real-world benchmarks cannot fully separate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes the first learning-based framework for arbitrary-scale RAW image downscaling. The method uses a wavelet-based recurrent reconstruction decoder (WRRD) that processes low-frequency and high-frequency wavelet sub-bands through two new modules: a Low-Frequency Arbitrary-Scale Downscaling Module (LASDM) and a High-Frequency Prediction Module (HFPM), together with an energy-maximization loss. The authors also introduce the Real-NIRD dataset of real HR-LR RAW pairs with a nominal 1.3x non-integer downscaling factor, and they evaluate on this dataset and on the Real-RawVSR dataset at 2x, 3x, and 4x. The central claim is state-of-the-art performance across all settings, with particularly large gains at 1.3x and 4x compared with interpolation baselines and learning-based sRGB downscaling methods.

Significance. If the empirical claims hold, this is a meaningful step for RAW-domain image downscaling: the task is new, the architecture is plausible, and the introduced Real-NIRD dataset plus the planned code release would be useful resources for the community. The paper gives credit to the lossless property of the wavelet transform and proposes a sensible coarse-to-fine recurrent design. However, the significance is currently conditional because several load-bearing aspects of the evaluation are not yet supported: the Real-NIRD alignment is unvalidated, the 1.3x comparison includes no learning-based baselines, hyperparameters appear to be selected on the evaluation split, and some numeric gains in the text do not match the tables. These issues do not invalidate the architecture, but they must be resolved before the state-of-the-art claim can be trusted.

major comments (4)
  1. [Sec. 4.1] The Real-NIRD dataset collection is described as using tripods and a remote trigger (Sec. 4.1), but no alignment validation is reported. Unlike Real-RawVSR, which used a beam splitter and RANSAC, the new dataset lacks any quantitative check for residual parallax, vibration, or lens distortion between the HR and LR pairs. Because the 1.3x state-of-the-art claim in Table 1 depends on the LR being a faithful physical 4/3 downscale of the HR scene, unvalidated misalignment of even 1-2 pixels could inflate the reported PSNR/SSIM/LPIPS gains. Please provide registration error statistics (e.g., residual homography error or feature correspondence distances) or employ a post-hoc alignment step, and describe the optical setup that guarantees identical field of view.
  2. [Sec. 4.3, Table 1] At the non-integer 1.3x setting, no learning-based baselines are reported (all entries are '-'), so the claim of state-of-the-art performance across all settings is supported only against interpolation methods. Since the main novelty includes non-integer downscaling, this is a critical omission. Please retrain and report at least AIDN [25] or IARN [16] at 1.3x, or explicitly qualify the 1.3x claim as being relative only to interpolation and RAW-based traditional methods.
  3. [Sec. 4.2 / Table 3] The ablation study in Table 3 appears to select the recurrent depth K and the loss weight lambda on the same evaluation split that is later used for the final reported numbers (Sec. 4.1 states that 637, 542, and 549 pairs are selected for evaluation). Choosing hyperparameters based on test performance biases the comparison in favor of the proposed configuration. In addition, no error bars or multiple-seed results are reported. Please fix the hyperparameters using a separate validation split and report mean and standard deviation over at least three training runs.
  4. [Sec. 4.3] The reported improvement numbers do not match Table 1. For the 4x setting, the text claims improvements of 0.53 PSNR, 0.016 SSIM, and 0.021 LPIPS, but the table shows differences relative to the best competitor of roughly 1.22 PSNR, 0.041 SSIM, and 0.054 LPIPS (vs. HCD-IRN). Similarly, for the 1.3x setting, the claimed 0.61 PSNR improvement is larger than the 0.49 gap between the method and the best baseline (Area_sRGB). Please correct the text or the table, and state which reference method is used for each gain.
minor comments (5)
  1. [Sec. 3.3, Eq. (6)] The loss is called "energy-maximization" but it actually minimizes the absolute difference between the L2 norms of the predicted and reference high-frequency coefficients. A name such as "energy-preservation" or "energy-consistency" would be more accurate and less confusing.
  2. [Abstract and Sec. 3.2] The phrase "information lossless attribute of wavelet transformation" is correct for the transform itself, but the overall downscaling process discards high-frequency information and predicts proxies for the target sub-bands, so the method is not information-lossless as a whole. Please rephrase to avoid implying perfect reconstruction.
  3. [Sec. 4.1] The non-integer factor is exactly 4/3 ≈ 1.333, not 1.3. Please state the exact rational factor or clarify that "1.3x" is an approximation.
  4. [Table 1] The caption states that second-best results are underlined, but no underlines appear in the table. Either add the underlines or remove the statement from the caption.
  5. [Sec. 4.3, Table 2] In the RAW-domain evaluation, only traditional interpolation methods are listed. Please clarify in the text that learning-based competitors are not compared in the RAW domain because they operate on sRGB, and note this as a limitation when interpreting the RAW-domain gains.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: supervised learning against physically captured LR references; wavelet 'lossless' framing is motivational, not definitional.

full rationale

No circular reasoning is present. The framework is a standard supervised pipeline: the HR input is transformed by a fixed 2D Haar wavelet transform, the network predicts wavelet coefficients and an LR RAW image, and the losses L_con, L_hwc, and L_em compare those predictions against externally supplied targets (the captured LR RAW image and the HR-side DWT coefficients). None of these targets is defined in terms of the network output, and no parameter is fitted to an evaluation subset and then reported as a prediction. The 'information lossless attribute of wavelet transformation' is used as design motivation for the coarse-to-fine decoder, but the learned reconstruction is explicitly supervised and is not claimed to inherit losslessness by construction; the overstatement, if any, is a soundness issue rather than circularity. The reference list contains no work by the present authors, so there is no self-citation chain and no imported uniqueness theorem. The strongest empirical claim, state-of-the-art performance at 1.3x on Real-NIRD, depends on the unvalidated pixel alignment of the captured HR-LR pairs, but that is an external-validity concern about whether the benchmark rewards registration artifacts; it is not a case of the paper's derivation reducing to its own inputs. Accordingly, no circular steps are identified and the score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The framework is a supervised learning system; the main uncharged assumptions are representational (wavelet coefficients can be predicted across scales, top-K HR locations map to LR locations) and a loss design choice (matching HF energy). No new physical entity is introduced. Hyperparameters K and lambda are tuned without a separate validation split, and the Real-NIRD ground truth alignment is not validated.

free parameters (2)
  • K (number of recurrent levels) = 4
    Chosen by ablation on the 2x Real-RawVSR evaluation split (Table 3). No separate validation set is used, so this is a free parameter tuned to the reported benchmark.
  • lambda (energy loss weight) = 0.1
    Hand-set in Sec 4.2 with no sensitivity analysis; the final numbers depend on it.
assumptions (4)
  • standard math Haar 2D-DWT and 2D-IDWT are information-lossless and invertible for packed RAW images.
    Invoked in Sec 3.2, Eqs. (1)-(2). The losslessness holds only if all coefficients are retained exactly, but the learned pipeline reconstructs coefficients approximately.
  • domain assumption A rational scale factor n/m can be implemented by pixel unshuffle with factor m followed by pixel shuffle with factor n, preserving low-frequency structure.
    Core of LASDM in Sec 3.3; assumes channel-merging ResBlocks preserve low-frequency information and that any fraction can be decomposed this way.
  • ad hoc to paper The high-frequency energy of the downscaled image should equal the high-frequency energy of the original image.
    Used in Eq. (6) as L_em. This is a design choice; normal downscaling typically reduces high-frequency energy, and matching it may induce aliasing or oversharpening.
  • domain assumption Top-K selection in the HR wavelet high-frequency maps identifies the coefficients that should appear in the LR wavelet maps.
    Used in HFPM, Sec 3.3; no proof that top-K positions in HR correspond to correct positions in LR after downscaling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Arbitrary-Scale RAW Image Downscaling with Wavelet-based Recurrent Reconstruction." pith.science (2026). https://pith.science/paper/7ZKBCOQ2

@misc{pith2026250723219,
  author       = {Pith},
  title        = {Pith review of: Learning Arbitrary-Scale RAW Image Downscaling with Wavelet-based Recurrent Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ZKBCOQ2}},
  note         = {Machine review of arXiv:2507.23219}
}
abstract

Image downscaling is critical for efficient storage and transmission of high-resolution (HR) images. Existing learning-based methods focus on performing downscaling within the sRGB domain, which typically suffers from blurred details and unexpected artifacts. RAW images, with their unprocessed photonic information, offer greater flexibility but lack specialized downscaling frameworks. In this paper, we propose a wavelet-based recurrent reconstruction framework that leverages the information lossless attribute of wavelet transformation to fulfill the arbitrary-scale RAW image downscaling in a coarse-to-fine manner, in which the Low-Frequency Arbitrary-Scale Downscaling Module (LASDM) and the High-Frequency Prediction Module (HFPM) are proposed to preserve structural and textural integrity of the reconstructed low-resolution (LR) RAW images, alongside an energy-maximization loss to align high-frequency energy between HR and LR domain. Furthermore, we introduce the Realistic Non-Integer RAW Downscaling (Real-NIRD) dataset, featuring a non-integer downscaling factor of 1.3$\times$, and incorporate it with publicly available datasets with integer factors (2$\times$, 3$\times$, 4$\times$) for comprehensive benchmarking arbitrary-scale image downscaling purposes. Extensive experiments demonstrate that our method outperforms existing state-of-the-art competitors both quantitatively and visually. The code and dataset will be released at https://github.com/RenYangSCU/ASRD.

Figures

Figures reproduced from arXiv: 2507.23219 by the authors.

Figure 1
Figure 1. Visual comparisons of the traditional interpolation-based image downscaling method Bicubic that conducted in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall pipeline of our proposed arbitrary-scale RAW image downscaling framework. Our approach first ex [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The detailed architecture of our proposed low [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The detailed architecture of our proposed high [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of our method with competitive methods Bilinear [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of our method with competitive methods CAR [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visual results of the ablation study about our recur [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Visual results of the ablation study about our newly [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages

  1. [25]

    Jinbo Xing, Wenbo Hu, Menghan Xia, and Tien-Tsin Wong. 2023. Scale-arbitrary invertible image downscaling. IEEE Transactions on Image Processing (2023)

  2. [16]

    Zhihong Pan, Baopu Li, Dongliang He, Wenhao Wu, and Errui Ding. 2023. Ef- fective invertible arbitrary image rescaling. In IEEE/CVF Winter Conference on Applications of Computer Vision. 5416–5425

  3. [1]

    Jingwei Bao, Jinhua Hao, Pengcheng Xu, Ming Sun, Chao Zhou, and Shuyuan Zhu. 2024. Plug-and-Play Tri-Branch Invertible Block for Image Rescaling. arXiv preprint arXiv:2412.13508 (2024)

  4. [2]

    Yinbo Chen, Sifei Liu, and Xiaolong Wang. 2021. Learning continuous image representation with local implicit image function. In IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8628–8638

  5. [3]

    Martin A Fischler and Robert C Bolles. 1981. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24, 6 (1981), 381–395

  6. [4]

    Eduardo SL Gastal and Manuel M Oliveira. 2017. Spectral remapping for image downscaling. ACM Transactions on Graphics 36, 4 (2017), 1–16

  7. [5]

    Alfred Haar. 1911. Zur theorie der orthogonalen funktionensysteme. Math. Ann. 71, 1 (1911), 38–53

  8. [6]

    Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. 2019. Meta-SR: A magnification-arbitrary network for super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1575–1584

Show all 31 references
  1. [7]

    Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. 2018. Task-aware image downscaling. In European Conference on Computer Vision . 399–414

  2. [8]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

  3. [9]

    Jaewon Lee and Kyong Hwan Jin. 2022. Local texture estimator for implicit representation function. In IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1929–1938

  4. [10]

    Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. 2018. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing 28, 3 (2018), 1092–1107

  5. [11]

    Yuanbang Liang, Bhavesh Garg, Paul Rosin, and Yipeng Qin. 2024. Deep genera- tive model based rate-distortion for image downscaling assessment. In IEEE/CVF Conference on Computer Vision and Pattern Recognition . 19363–19372

  6. [12]

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. 2017. Enhanced deep residual networks for single image super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops . 136–144

  7. [13]

    2017.𝐿_{0}-regularized image downscaling

    Junjie Liu, Shengfeng He, and Rynson WH Lau. 2017.𝐿_{0}-regularized image downscaling. IEEE Transactions on Image Processing 27, 3 (2017), 1076–1085

  8. [14]

    Don P Mitchell and Arun N Netravali. 1988. Reconstruction filters in computer- graphics. SIGGRAPH 22, 4 (1988), 221–228

  9. [15]

    A Cengiz Oeztireli and Markus Gross. 2015. Perceptually based downscaling of images. ACM Transactions on Graphics 34, 4 (2015), 1–10

  10. [17]

    Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. 2016. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE/CVF Conference on Computer Vision ...

  11. [18]

    Sanghyun Son and Kyoung Mu Lee. 2021. Srwarp: Generalized image super- resolution under arbitrary transformation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition . 7782–7791

  12. [19]

    Wanjie Sun and Zhenzhong Chen. 2020. Learned image downscaling for upscaling using content adaptive resampler.IEEE Transactions on Image Processing 29 (2020), 4027–4040

  13. [20]

    Wanjie Sun and Zhenzhong Chen. 2024. Learning many-to-many mapping for unpaired real-world image super-resolution and downscaling. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

  14. [21]

    Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. 2021. Learning a single network for scale-arbitrary super-resolution. In IEEE/CVF International Conference on Computer Vision . 4801–4810

  15. [22]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing 13, 4 (2004), 600–612

  16. [23]

    Mingqing Xiao, Shuxin Zheng, Chang Liu, Zhouchen Lin, and Tie-Yan Liu. 2023. Invertible rescaling network and its extensions.International Journal of Computer Vision 131, 1 (2023), 134–159

  17. [24]

    Mingqing Xiao, Shuxin Zheng, Chang Liu, Yaolong Wang, Di He, Guolin Ke, Jiang Bian, Zhouchen Lin, and Tie-Yan Liu. 2020. Invertible image rescaling. In European Conference on Computer Vision . 126–144

  18. [26]

    Bingna Xu, Yong Guo, Luoqian Jiang, Mianjie Yu, and Jian Chen. 2023. Down- scaled representation matters: Improving image rescaling with collaborative downscaled images. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 12237–12247

  19. [27]

    Jie-En Yao, Li-Yuan Tsao, Yi-Chen Lo, Roy Tseng, Chia-Che Chang, and Chun- Yi Lee. 2023. Local implicit normalizing flow for arbitrary-scale image super- resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1776–1785

  20. [28]

    Huanjing Yue, Zhiming Zhang, and Jingyu Yang. 2022. Real-rawvsr: Real-world raw video super-resolution with a benchmark dataset. In European Conference on Computer Vision. 608–624

  21. [29]

    Wengyi Zhan, Mingbao Lin, Chia-Wen Lin, and Rongrong Ji. 2024. AnySR: Realizing Image Super-Resolution as Any-Scale, Any-Resource.IEEE Transactions on Image Processing (2024)

  22. [30]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  23. [2018]

    In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition . 586–595

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.