{"id":"7e70f2cc-818d-4144-80d6-f2bd34d1b144","arxiv_id":"2507.23219","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A wavelet-based recurrent network performs arbitrary-scale downscaling directly on RAW images, outperforming sRGB and interpolation baselines on new and existing benchmarks.","lead":"This paper presents a learning-based method for downscaling RAW images to any size, using wavelets to split the image into low- and high-frequency parts and reconstruct the smaller image level by level. It also introduces a new dataset of real RAW photo pairs for non-integer downscaling and reports better quality than existing downscaling methods on several benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-NIRD alignment is unvalidated; the 1.3x SOTA claim may largely reflect reference misalignment, and the RAW-domain metric for learning-based sRGB methods is missing.","rationale":"The weakest assumption is the pixel alignment of Real-NIRD, and the reader's weakest_assumption identifies exactly this. The proposed test directly measures the residual alignment and checks whether the quantitative gains survive on well-aligned pairs. A secondary but related issue is the missing RAW-domain evaluation for all learning-based methods, which is also addressed by the same conditional recommendation: if the concern lands, the paper should either add alignment error bars or downgrade the 1.3x SOTA claim. The verdict remains CONDITIONAL because the concern is empirical and addressable; it does not invalidate the framework, the component ablations, or the integer-scale results on Real-RawVSR (which has documented alignment). No change to the reader's verdict is needed, so verdict_should_be is CONDITIONAL (unchanged from reader).","tokens_in":13657,"tokens_out":1277,"duration_ms":12408,"concrete_test":"For all 149 Real-NIRD test pairs, estimate per-pair residual motion between HR and LR (e.g., phase correlation or ECC on the RAW Bayer pattern before demosaic, fitting translation/rotation/scale). Report the distribution of residuals and re-evaluate all methods on a subset with residual < 0.25 px. If the 1.3x gains shrink or vanish on the well-aligned subset, the SOTA claim at 1.3x must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim of SOTA at 1.3x rests on the Real-NIRD HR-LR pairs being pixel-aligned physical 4/3 downscales (Sec 4.1). The paper only says 'tripods' and 'remote applications' were used; no beam splitter, no RANSAC, no alignment error analysis, unlike the Real-RawVSR description. Uncorrected parallax, vibration, or slight zoom differences can shift edges by 1-2 px, which directly inflates PSNR/SSIM/LPIPS for any learned method that can reproduce a slight registration warp. While Table 2 gives RAW-domain PSNR/SSIM, it only compares interpolation methods; the learning-based sRGB baselines (CAR, IRN, AIDN, TIRN) are evaluated only after ISP conversion, so their RAW-domain performance is not established. The interpolation baselines at 1.3x are themselves computed from mismatched LR references, so the 0.61 PSNR / 0.055 SSIM / 0.066 LPIPS gains at 1.3x (Sec 4.3) are not yet attributable to downscaling quality rather than alignment. This does not disprove the method, but it makes the strongest empirical claim conditional until registration is checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":13919,"tokens_out":8468,"duration_ms":87713,"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":[{"comment":"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.","section":"Sec. 4.1"},{"comment":"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.","section":"Sec. 4.3, Table 1"},{"comment":"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.","section":"Sec. 4.2 / Table 3"},{"comment":"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.","section":"Sec. 4.3"}],"minor_comments":[{"comment":"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.","section":"Sec. 3.3, Eq. (6)"},{"comment":"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.","section":"Abstract and Sec. 3.2"},{"comment":"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.","section":"Sec. 4.1"},{"comment":"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.","section":"Table 1"},{"comment":"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.","section":"Sec. 4.3, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a potentially useful architecture and a valuable new dataset, but the central state-of-the-art claim is not yet fully supported because of the unvalidated Real-NIRD alignment, the absence of learning-based baselines at 1.3x, the hyperparameter selection on the evaluation split, and the numeric inconsistencies in Sec. 4.3. I would support acceptance after these issues are addressed with additional experiments and clarifications."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper introduces a genuinely new task: learning-based arbitrary-scale RAW image downscaling, with a new non-integer benchmark dataset (Real-NIRD). That alone is a useful contribution for RAW storage and computational photography. The wavelet-based recurrent coarse-to-fine design is a reasonable combination, and the ablations show each component helps. I agree with the reader that there is no circularity problem; the LR references are captured images, not model outputs.\n\nThe soft spots are empirical, not fatal. The biggest issue is the Real-NIRD alignment. The paper says only that tripods and a remote trigger were used; there is no registration, no RANSAC, no alignment-error analysis. The Real-RawVSR pairs went through a beam splitter plus RANSAC, so those integer-scale results are more trustworthy. But the 1.3x SOTA claim (Table 1) heavily depends on the 149 evaluation pairs being pixel-aligned. If edges are shifted by 1-2 px, PSNR/SSIM/LPIPS gains partly reflect registration warps rather than downscaling quality. The stress-test note is correct to flag this; it does not disprove the method but makes the headline claim conditional until the authors show alignment checks.\n\nTwo smaller empirical concerns: no error bars, and K and lambda are tuned on the same split used for final numbers (Table 3 is on the 2x eval split). Also, the learning-based sRGB baselines are not evaluated in the RAW domain, only after ISP conversion, so Table 2 compares only interpolation methods. The \"information lossless\" phrase for wavelet transforms is an overstatement; the network learns a lossy mapping, even if the transform itself is invertible.\n\nThe paper is clearly written, the experiments are reasonably thorough, and the dataset is a useful resource. If the alignment issue is addressed, the central claims likely hold for the integer scales. I would bring this to a reading group interested in computational photography, and I would cite it if I worked on RAW image processing. It deserves a serious peer review; the concerns are addressable in revision.","headline":"First learning-based arbitrary-scale RAW downscaling with a sensible wavelet design; the 1.3x SOTA claim is conditional on unvalidated ground-truth alignment.","tokens_in":14450,"tokens_out":1314,"would_cite":true,"duration_ms":15857,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["RAW image downscaling","arbitrary-scale downscaling","wavelet transform","recurrent reconstruction","high-frequency prediction","Real-NIRD dataset","energy-maximization loss","image rescaling"],"falsifier":"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.","tokens_in":13446,"feed_emoji":"📷","tokens_out":5913,"duration_ms":53779,"temperature":0.7,"pith_summary":"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.","feed_headline":"RAW downscaling at any scale, even 1.3x, keeps sharp detail","feed_subtitle":"A wavelet recurrent network beats interpolation and sRGB rescaling on real RAW photos, backed by a new non-integer benchmark.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the Real-RawVSR dataset with integer-scale RAW LR-HR pairs used for training and evaluation.","marker":"[28]"},{"why":"Serves as a baseline learning-based sRGB image downscaling method that the proposed approach must outperform.","marker":"[24]"},{"why":"Provides the AIDN method, a baseline for arbitrary-scale sRGB downscaling.","marker":"[25]"},{"why":"Provides the TIRN method, a baseline tri-branch invertible image rescaling approach.","marker":"[1]"},{"why":"Provides the CAR method, a baseline content-adaptive resampler for downscaling.","marker":"[19]"},{"why":"Provides the HCD-IRN method, a baseline collaborative downscaled-image approach.","marker":"[26]"},{"why":"Supplies the Haar wavelet basis used in the 2D discrete wavelet transform that the method builds on.","marker":"[5]"},{"why":"Supplies the pixel shuffle and pixel unshuffle operations that the LASDM uses for arbitrary-scale channel merging.","marker":"[17]"},{"why":"Supplies the RANSAC alignment algorithm used for the Real-RawVSR pairs, which is the basis for assuming those pairs are correctly aligned.","marker":"[3]"}],"fun_headline_variants":["Wavelet RAW downscaling hits any scale, even 1.3x","Arbitrary-scale RAW downscaling via wavelet recurrence","Wavelet-based RAW downscaler: any scale, sharp detail","RAW downscaling at arbitrary scales, even non-integer","First learning-based arbitrary-scale RAW downscaler"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet RAW downscaling hits any scale, even 1.3x","Arbitrary-scale RAW downscaling via wavelet recurrence","Wavelet-based RAW downscaler: any scale, sharp detail","RAW downscaling at arbitrary scales, even non-integer","First learning-based arbitrary-scale RAW downscaler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001121,"raw_usage":{"total_tokens":4702,"prompt_tokens":1017,"completion_tokens":3685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":3600}},"tokens_in":633,"tokens_out":3685,"duration_ms":27058,"temperature":1.0,"reasoning_tokens":3600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:56:53.206603+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Real-RawVSR dataset with integer-scale RAW LR-HR pairs used for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a baseline learning-based sRGB image downscaling method that the proposed approach must outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the AIDN method, a baseline for arbitrary-scale sRGB downscaling."},{"cited_title":"Plug-and-Play Tri-Branch Invertible Block for Image Rescaling","cited_arxiv_id":"2412.13508","evidence_quote":"Provides the TIRN method, a baseline tri-branch invertible image rescaling approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CAR method, a baseline content-adaptive resampler for downscaling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HCD-IRN method, a baseline collaborative downscaled-image approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Haar wavelet basis used in the 2D discrete wavelet transform that the method builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pixel shuffle and pixel unshuffle operations that the LASDM uses for arbitrary-scale channel merging."}],"review_version":1}