{"id":"5168a3a5-8b04-44e7-b555-8e2770223ff8","arxiv_id":"2506.03645","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single synthetic-trained SNR-guided denoiser, wrapped in coarse-to-fine noise estimation and expectation-matched variance stabilization, denoises raw images from unseen cameras without camera-specific training.","lead":"YOND is a raw image denoising method that trains one denoiser on synthetic noise and then works on any camera's raw images without recalibration. It combines per-image noise estimation, a variance-stabilizing transform with bias correction, and an SNR-guided neural denoiser.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The σSNR guidance mapping in §III-D is internally inconsistent with the unit-variance VST output and with SNR-Net's training range; the actual scaling fed to SNR-Net is undocumented, so the central blind-deployment claim depends on an unstated normalization step.","rationale":"The reader's weakest assumption correctly identifies the σSNR mapping as the critical unverified step, and my analysis confirms it is the most load-bearing concern. The paper's own text contains two incompatible characterizations: a theoretical reciprocal-of-SNR definition that would put σSNR below the training range, and a real-world example with σSNR=32.87 that is within the training range but contradicts the reciprocal formula and the unit-variance VST claim. The gap is not merely cosmetic—if the denoiser actually receives a unit-variance transformed image while conditioned on a raw-domain noise-level value, it is being used outside its training distribution, and the reported strong generalization would require an undocumented scale transformation. Because the code is not yet available, this missing implementation detail blocks verification of the central claim. The paper has substantial independent support in its extensive experiments, ablations, and honest failure-mode analysis, so the appropriate verdict remains CONDITIONAL (acceptance contingent on clarifying the σSNR computation and input scaling), matching the reader's assessment. I therefore recommend UNCHANGED rather than ACCEPT or REJECT, as the concern is substantive but not yet disproven by evidence in the manuscript.","tokens_in":31431,"tokens_out":11223,"duration_ms":110789,"concrete_test":"Reproduce the pipeline on a synthetic noisy raw image with known (α,σ): compute the EM-VST output f(y), measure the empirical noise std of f(y)−f(x) (should be ≈1), then compute σSNR using the §III-D reciprocal-of-SNR formula and also estimate the raw-domain noise std. Run SNR-Net with three conditioning values: (a) the reciprocal-of-SNR value, (b) the actual transformed-domain std (≈1), and (c) the raw-domain std (e.g., ~32.87). If (a) or (b) underperforms (c) by more than ~1 dB PSNR, or if the code reveals an undocumented input scaling, the paper's described σSNR mapping is incorrect and the central generalization claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-D states that after VST the noise variance is approximately 1, then defines σSNR as the reciprocal of the SNR, i.e., 1/peak of the transformed image. For typical VST outputs with peak values of 10–100, this yields σSNR≈0.01–0.1. Yet SNR-Net was trained on raw-domain AWGN with σ∈[5,50] (Section IV-A1), and Section V-A3 reports a concrete instance with σSNR=32.87 'after VST'—a value within [5,50] but unreachable from the reciprocal-of-SNR rule (it would require a transformed peak of ≈0.03) and incompatible with a unit-variance VST. The paper never specifies whether the VST output is rescaled before being fed to SNR-Net, how σSNR is computed in actual code, or how the conditioning is normalized to match the training distribution. If the input is truly f(y) with noise std≈1 and conditioning σSNR=32.87, SNR-Net operates far outside its training range; the only way the reported results can hold is an undocumented scaling/normalization of the transformed image or of σSNR. This is load-bearing because the entire blind-deployment claim rests on SNR-Net correctly processing VST-transformed data under the stated guidance mapping.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes YOND, a blind raw-image denoising pipeline that combines coarse-to-fine noise estimation (CNE), an expectation-matched variance-stabilizing transform (EM-VST), and a controllable AWGN denoiser (SNR-Net) trained once on synthetic data. The pipeline estimates per-image Poisson-Gaussian noise parameters, transforms the noisy raw image to approximately Gaussian noise, corrects the VST bias before denoising, and feeds the transformed image and an SNR-derived noise level to a lightweight UNet. The authors report superior PSNR/SSIM over camera-specific calibration-based and self-supervised baselines on ELD, LRID, SIDD, and DND datasets, and they propose an optional DDIM-based iterative extension (YOND-p) for generative detail synthesis. The central claim is that YOND generalizes to unseen cameras without camera-specific training data.","tokens_in":31729,"tokens_out":7817,"duration_ms":82950,"significance":"If the result holds, YOND would be a practically important step toward camera-agnostic raw denoising. The method is evaluated on multiple external benchmarks with per-image noise estimation, which avoids the circularity of fitting noise parameters to the test set, and the paper explicitly discusses challenging cases and model limitations. The proposed EM-VST bias correction is a sensible and well-motivated improvement over UIAT for low-light regions, and the pipeline-level design makes the method amenable to manual control. The main weakness is that the mapping from the VST output to the conditioning input of SNR-Net is unspecified, which currently prevents reproduction and, if taken literally, contradicts both the unit-variance assumption and the network's training range. With a precise specification of this scaling, the central claims would be substantially supported.","major_comments":[{"comment":"The definition of σ_SNR is internally inconsistent and the actual conditioning procedure is not described. Section III-D states that after VST the noise variance is approximately 1 and defines σ_SNR as the reciprocal of the SNR, i.e., 1/peak of the transformed image. For typical VST outputs with peak values of 10–100, this yields σ_SNR ≈ 0.01–0.1, far below the training range σ∈[5,50] reported in Section IV-A1. Yet Section V-A3 reports a concrete instance with σ_SNR=32.87 'after VST'—a value unreachable from the reciprocal-of-SNR rule (it would require a transformed peak of ≈0.03) and incompatible with a unit-variance VST. The paper never specifies whether the VST output is rescaled before being fed to SNR-Net, how σ_SNR is computed in the actual code, or how the conditioning is normalized to match the training distribution. Since the entire blind-deployment claim rests on SNR-Net correctly processing VST-transformed data under the stated guidance mapping, this missing normalization/scaling must be specified and validated, e.g., by reporting the distribution of σ_SNR values on the evaluated datasets.","section":"Section III-D, Algorithm 1, Section V-A3"},{"comment":"The paper reports that YOND achieves state-of-the-art results on 68.44% of SIDD patches and acknowledges 'noticeable degradation' in texture-rich scenarios, but does not report the average PSNR/SSIM on the remaining 31.56% or how the baselines behave on those patches. Because the abstract claims robust generalization, a quantitative breakdown of the failure cases (e.g., comparison of the worst-decile performance against baselines) would help the reader calibrate the scope of the central claim.","section":"Section V-A1"}],"minor_comments":[{"comment":"In the paragraph beginning 'Next, we will we introduce', the duplicated word 'we' should be removed for readability.","section":"Section V-A1"},{"comment":"The sentence 'The learning rate will vary with each epoch in a cosine annealing manner' is grammatically awkward; consider rewording to 'The learning rate follows a cosine annealing schedule across epochs.'","section":"Section IV-A1"},{"comment":"The units 'e-' and 'read noise (e-)' are not defined, and the horizontal axis is logarithmic but this is not stated in the caption; please clarify the axis labels and units.","section":"Figure 6 caption"},{"comment":"The parenthetical percentages represent deviations from the true noise parameters, but this is not stated; please add a note to the caption explaining the notation.","section":"Table V caption"}],"recommendation":"major_revision","confidential_remarks":"The primary obstacle is the unspecified σ_SNR computation and normalization of the VST output before it is fed to SNR-Net. The reported σ_SNR=32.87 in Section V-A3 strongly suggests that the formula in Section III-D is not what is actually implemented. I recommend requiring the authors to provide the exact computation, ideally with code, before making a final decision. If they cannot supply a coherent mapping, the blind-deployment claim would be unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Feng et al. have a real result here: a blind raw denoiser trained only on synthetic AWGN that transfers to unseen cameras, with three mechanisms—coarse-to-fine noise estimation, expectation-matched VST bias pre-correction, and SNR-guided conditioning—that are new enough and ablated well enough to matter. The evaluation is broad (ELD, LRID, SIDD, DND) and the method beats camera-specific calibration baselines on many subsets. The failure-mode discussion on flat-region-poor images is honest. Credit where due: this is a solid empirical paper.\n\nThe soft spot the reader flagged is real and load-bearing. Section III-D says the VST output has unit noise variance and defines σSNR as the reciprocal of the SNR, i.e., 1/peak of the transformed image. For typical VST outputs with peak values of tens, that gives σSNR around 0.01–0.1. Yet SNR-Net was trained on raw-domain AWGN with σ∈[5,50], and Section V-A3 reports σSNR=32.87 \"after VST\". A unit-variance image cannot have a 32.87 noise level. Somewhere between EM-VST and SNR-Net there is an undocumented rescaling of the transformed image or of σSNR itself. The paper never specifies this normalization, and without it the reported results are hard to reconcile with the stated training range. This is exactly the kind of implementation detail that would break blind deployment if the recipe were followed naively. It is fixable in a revision, but the authors need to state the exact computation and any scaling they apply.\n\nMinor points: source code is promised but not available, so external reproducibility is pending. The paper's self-citations are to prior work in the same group (FBI-Denoiser, PMN, DMID) and are not a problem in themselves. The σSNR confusion is the main technical gap.\n\nOverall: the central claim—blind generalization from synthetic training—is plausible and mostly supported by the experiments, but the missing normalization step in the conditioning path prevents full verification. This paper deserves a serious referee and can likely be accepted after a major revision clarifying the pipeline. I would not desk reject it.","headline":"Solid blind raw denoiser with genuinely useful modules, but the σSNR guidance mapping is under-specified and internally inconsistent; the results are plausible but the paper needs a major revision to document the actual scaling.","tokens_in":32238,"tokens_out":4165,"would_cite":true,"duration_ms":46406,"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":"Trained only on synthetic noise, YOND generalizes to raw images from unknown cameras by combining coarse-to-fine noise estimation, expectation-matched variance stabilization, and an SNR-guided denoiser.","keywords":["blind raw image denoising","camera-specific data dependency","Poisson-Gaussian noise","variance-stabilizing transform","coarse-to-fine noise estimation","SNR-guided denoiser","synthetic training data","generative denoising"],"falsifier":"Synthesize raw images with known Poisson-Gaussian parameters across a wide signal range, run CNE+EM-VST, and measure the variance of transformed flat patches; if it deviates from 1, or if the estimated $\\sigma_{\\mathrm{SNR}}$ disagrees with the reciprocal of the measured SNR, the asserted reduction to AWGN fails and with it the blind generalization.","tokens_in":31246,"feed_emoji":"📷","tokens_out":8020,"duration_ms":82322,"temperature":0.7,"pith_summary":"This paper introduces YOND, a blind raw-image denoising pipeline that is trained once on synthetic additive white Gaussian noise and then applied to raw images from arbitrary cameras without any per-camera calibration or paired data. The method decomposes the problem into three modules: coarse-to-fine noise estimation (CNE) that recovers a camera's Poisson-Gaussian noise parameters from the noisy image alone; an expectation-matched variance-stabilizing transform (EM-VST) that pre-corrects the bias of the Anscombe transform; and an SNR-guided denoiser (SNR-Net) that denoises the stabilized image under explicit noise-level control. On the ELD, LRID, SIDD, and DND datasets, the paper reports that this one-time-trained pipeline outperforms or matches methods that had camera-specific calibration or self-supervised training data. If true, the practical bottleneck for raw denoising shifts from data collection to single-image noise estimation.","feed_headline":"One denoiser, any camera: raw noise gone after one training","feed_subtitle":"YOND trains once on synthetic noise, then estimates and stabilizes each camera's noise to denoise raw images it has never seen.","key_machinery":"The load-bearing object is the reduction of arbitrary camera noise to a standard AWGN denoising problem. CNE fits $V=\\alpha I+\\sigma^2$ by least squares over flat-region masks, first on the noisy image and then on the noisy image with the variance of the coarse denoised image subtracted, giving the fine estimate $(\\alpha_f,\\sigma_f)$. EM-VST applies $f_{\\hat\\sigma}(z)=2\\sqrt{z+3/8+\\hat\\sigma^2}$ and subtracts the bias function $e_{\\hat\\sigma}(x)=E(f_{\\hat\\sigma}(z)\\mid x)-f_{\\hat\\sigma}(x)$ evaluated on the noisy image, so bias is corrected before denoising. SNR-Net is a UNet with residual SNR-Blocks that take the noise level $\\sigma_{\\mathrm{SNR}}$; because the stabilized noise has unit variance, $\\sigma_{\\mathrm{SNR}}$ is the reciprocal of the peak-to-noise ratio, giving the network an explicit dial for denoising strength. A DDIM-style iteration over SNR-Net (YOND-p) extends the same weights to generate detail when perceptual quality is wanted.","core_discovery":"The central claim is that camera-specific data dependency in raw denoising can be broken by a 'denoise after you stabilize' strategy. YOND assumes each camera's noise is Poisson-Gaussian, estimates the two parameters ($\\alpha$ and $\\sigma$) from flat regions of the noisy image, then refines that estimate by subtracting the variance of a coarse denoised version; it then applies a generalized Anscombe transform whose expectation bias is corrected against the noisy image itself; and finally it feeds the resulting near-AWGN image, plus a noise level equal to the reciprocal of its SNR, to a UNet-style denoiser trained once on synthetic AWGN. The paper argues this removes the need for camera-specific training, and supports the claim with blind results on unknown cameras from ELD, LRID, SIDD, and DND that are higher in PSNR than calibration-based and self-supervised baselines in most tested configurations.","pith_inferences":["Editorial inference: CNE plus EM-VST could be packaged as a camera-agnostic front end for any pretrained AWGN denoiser, making the blind-generalization claim testable independently of SNR-Net.","Editorial inference: The same expectation-matching trick should transfer to other variance-stabilizing transformations with smooth bias functions, which would widen YOND's scope beyond Poisson-Gaussian noise while keeping the pre-correction logic.","Editorial inference: The failure mode the paper identifies - texture-rich scenes where flat-region masks are unreliable - points to uncertainty-aware noise estimation as the most promising next step; manual $\\sigma_{\\mathrm{SNR}}$ adjustment handles the failure but leaves automation open."],"forward_implications":["A single training run on synthetic AWGN suffices for deployment on cameras never seen in training, removing the standard per-camera calibration and paired-data collection steps.","Users can trade noise removal against detail by adjusting $\\sigma_{\\mathrm{SNR}}$, so the same network serves multiple noise levels instead of needing one model per level.","Because bias correction happens before denoising, EM-VST prevents low-light color bias even when the downstream denoiser is a diffusion model, not just a regressive UNet.","The coarse-to-fine refinement improves noise parameter accuracy over single-step estimators, which the ablations show transfers into higher downstream PSNR on SIDD."],"supporting_citations":[{"why":"Supplies the Poisson-Gaussian noise model $y\\sim\\alpha P(x/\\alpha)+N(0,\\sigma^2)$ and the single-image estimation baseline that CNE refines.","marker":"[18]"},{"why":"Defines the generalized Anscombe transform that EM-VST stabilizes and debiases.","marker":"[26]"},{"why":"Provides the exact unbiased inverse of the Anscombe transform and the closed-form approximation used for high-signal regions.","marker":"[30]"},{"why":"Demonstrates the VST-plus-denoising pipeline and the learned noise estimator PGE that CNE compares against and improves.","marker":"[21]"},{"why":"Provides the unprocessing technique that converts sRGB images into pseudo raw images for training SNR-Net on camera-independent synthetic data.","marker":"[3]"},{"why":"Supplies the DDIM update rule used in the iterative YOND-p extension for detail generation.","marker":"[36]"},{"why":"Provides the SIDD real-world smartphone raw dataset used to test cross-camera generalization.","marker":"[14]"}],"fun_headline_variants":["YOND: one denoiser, all cameras, no retraining","Blind raw denoising without camera-specific data","Synthetic-only training works on unknown camera raw noise","YOND: estimate noise, stabilize, then denoise any camera raw","Break camera dependency: denoise with one synthetic-trained model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an arbitrary camera's noise is Poisson-Gaussian and that after CNE and EM-VST it becomes additive white Gaussian noise whose standard deviation is the reciprocal of the SNR, and that this value lies in the range SNR-Net was trained on; if any part fails, the blind generalization claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["YOND: one denoiser, all cameras, no retraining","Blind raw denoising without camera-specific data","Synthetic-only training works on unknown camera raw noise","YOND: estimate noise, stabilize, then denoise any camera raw","Break camera dependency: denoise with one synthetic-trained model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000463,"raw_usage":{"total_tokens":2345,"prompt_tokens":1004,"completion_tokens":1341,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1256}},"tokens_in":620,"tokens_out":1341,"duration_ms":10330,"temperature":1.0,"reasoning_tokens":1256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:58:27.199799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize raw images with known Poisson-Gaussian parameters across a wide signal range, run CNE+EM-VST, and measure the variance of transformed flat patches; if it deviates from 1, or if the estimated $\\sigma_{\\mathrm{SNR}}$ disagrees with the reciprocal of the measured SNR, the asserted reduction to AWGN fails and with it the blind generalization.","supporting_citations":[{"cited_title":"Practical poissonian-gaussian noise modeling and fitting for single-image raw- data,","cited_arxiv_id":null,"evidence_quote":"Supplies the Poisson-Gaussian noise model $y\\sim\\alpha P(x/\\alpha)+N(0,\\sigma^2)$ and the single-image estimation baseline that CNE refines."},{"cited_title":"Starck, F","cited_arxiv_id":null,"evidence_quote":"Defines the generalized Anscombe transform that EM-VST stabilizes and debiases."},{"cited_title":"Optimal inversion of the generalized anscombe transformation for poisson-gaussian noise,","cited_arxiv_id":null,"evidence_quote":"Provides the exact unbiased inverse of the Anscombe transform and the closed-form approximation used for high-signal regions."},{"cited_title":"Fbi-denoiser: Fast blind image denoiser for poisson-gaussian noise,","cited_arxiv_id":null,"evidence_quote":"Demonstrates the VST-plus-denoising pipeline and the learned noise estimator PGE that CNE compares against and improves."},{"cited_title":"Unprocessing images for learned raw denoising,","cited_arxiv_id":null,"evidence_quote":"Provides the unprocessing technique that converts sRGB images into pseudo raw images for training SNR-Net on camera-independent synthetic data."},{"cited_title":"Denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Supplies the DDIM update rule used in the iterative YOND-p extension for detail generation."},{"cited_title":"A high-quality denoising dataset for smartphone cameras,","cited_arxiv_id":null,"evidence_quote":"Provides the SIDD real-world smartphone raw dataset used to test cross-camera generalization."}],"review_version":1}