{"id":"467a8d3c-cc33-4319-9a51-bd5ca27b2050","arxiv_id":"2507.01587","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Camera metadata plugged into a denoising network via adaptive layer normalization adds user control and modest quality gains, but lacks a parameter-matched baseline.","lead":"This paper trains a denoising network to use camera settings (ISO, shutter speed, aperture) as extra inputs, letting users control how much noise is removed. The method improves quality on one dataset but barely changes results on another, and the gains are not clearly separated from extra model capacity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Camera-conditioning benefit is confounded with added parameters: CPADNet has +42% params versus baseline, only SID shows a real gain, and no capacity-matched or condition-ablated control is run.","rationale":"The reader's weakest assumption is exactly the capacity confound, and I agree it is the most load-bearing gap in the paper's argument. The proposed novelty is camera-parameter conditioning, but the only controlled comparison is baseline NAFNet versus CPADNet, where both the architecture (adaLN) and the parameter count change. Without a capacity-matched baseline or a condition-ablated variant, the SID improvement cannot be attributed to the camera-parameter prior. The SIDD results reinforce the concern: a 42% parameter increase buys only about 0.02 dB on SIDD, so the larger SID gain may be dataset-specific and driven by extra parameters rather than by the conditioning mechanism. The paper's own discussion of ISO distribution differences between SID and SIDD does not rule out this explanation. Controllability is demonstrated only qualitatively, and while that is suggestive, it does not address the quantitative attribution problem. The reader's conditional verdict, requiring a capacity-matched ablation and error bars, is appropriate; my stress-test does not identify a further objection that would move the verdict.","tokens_in":7960,"tokens_out":6257,"duration_ms":83018,"concrete_test":"Retrain the baseline NAFNet with channel width increased so that its parameter count matches CPADNet (9.45M), using the identical SID patch-crop protocol, L1 loss, 200K iterations, and evaluation setup, across three random seeds, and also report SIDD validation PSNR. If the matched-capacity baseline reaches roughly 25.5 dB on SID Sony patch PSNR, the reported +0.99 dB gain is capacity, not camera conditioning; if it stays near 24.6 dB, the conditioning mechanism is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that camera metadata conditioning improves state-of-the-art denoising and adds controllability. The main quantitative evidence is Table 1: CPADNet raises SID Sony patch PSNR from 24.58 to 25.57 (+0.99 dB) relative to the baseline, but CPADNet also has 9.45M parameters versus 6.67M for the baseline, a 42% increase. On SIDD, the same comparison yields only +0.02 dB PSNR, which is within ordinary run-to-run variation for a single training run. The paper acknowledges the parameter increase in Sec. 3.3 but provides no capacity-matched baseline and no ablation that removes or scrambles the camera-parameter information while keeping the architecture and parameter count fixed. Consequently, the +0.99 dB SID improvement could be explained entirely by the extra capacity of the adaLN conditioning MLP, rather than by the camera-parameter prior. The qualitative controllability demonstrations in Figs. 4 and 5 do not resolve this confound, because they show only that the network responds to its conditioning input, not that the performance gain is caused by meaningful camera metadata.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CPADNet, a denoising network that conditions a NAFBlock-based U-Net on camera parameters (ISO, shutter speed, F-number) via adaptive layer normalization. The camera parameters are non-linearly equalized into a 27-dimensional vector, and a learned embedding is used for smartphone cameras that lack F-number metadata. Experiments on the SID Sony subset and SIDD compare CPADNet with the baseline NAFNet and CPA-Restormer with Restormer, reporting PSNR/SSIM gains and qualitative inference-time controllability demonstrations in which the user changes the conditioning vector to adjust denoising strength.","tokens_in":8208,"tokens_out":3862,"duration_ms":47873,"significance":"If the reported performance gains are attributable to the camera-parameter conditioning, the method is practically attractive: it uses ground-truth EXIF metadata rather than estimated noise levels, is end-to-end trainable, adds almost no MACs, and is presented as a plug-in module for existing restoration networks. The code is promised to be released. However, the empirical evidence is currently insufficient to separate the effect of the camera-conditioning prior from the effect of added model capacity, and no comparison with prior noise-level-conditioning methods is provided. The core idea is plausible and worth pursuing, but the central quantitative claim needs stronger experimental support.","major_comments":[{"comment":"The central quantitative claim is confounded by model capacity. CPADNet has 9.45M parameters versus 6.67M for the baseline, a 42% increase, while MACs are identical (18.53G). On SIDD valid and test, the gain is only +0.02 dB (39.41 to 39.43 and 39.57 to 39.59), which is within ordinary run-to-run variation for a single training run. The paper acknowledges the parameter increase but provides no capacity-matched baseline, no ablation that removes or scrambles the camera-parameter information while keeping the architecture fixed, and no repeated-seed statistics. The same issue affects CPA-Restormer (11.72M to 15.06M parameters). Please add a capacity-matched baseline (e.g., a wider NAFNet with the same parameter count as CPADNet) and a condition-ablated control (e.g., a constant or shuffled camera vector), and report mean and standard deviation over at least three training runs.","section":"Section 3.3, Table 1"},{"comment":"The claim that the method improves state-of-the-art denoising networks is not supported by comparison with prior noise-level-conditioning or controllable denoising methods. CBDNet, DUDN, and variational approaches [3]-[6] also provide inference-time control of denoising strength, but no quantitative comparison is given on the same datasets. Without such baselines, the paper should either add a comparison table or delimit the claim to 'adds controllability via camera metadata' rather than 'improves state-of-the-art denoising networks.'","section":"Section 1 and Section 3.3"},{"comment":"The controllability demonstration is qualitative only, and some of the manipulated camera parameters may lie outside the training distribution (e.g., ISO 50 and F/22 for SID, and ISO 10000 for SIDD, whose training distribution in Fig. 3 has little support at such values). If these extreme conditioning vectors are extrapolations, the visual changes may reflect out-of-distribution behavior rather than a learned, meaningful denoising-strength axis. Please provide a quantitative evaluation of controllability, such as PSNR/SSIM or a noise-reduction metric as a function of the conditioning vector on held-out data, and clarify whether the shown parameter combinations occur in the training set.","section":"Section 3.4, Figs. 4 and 5"}],"minor_comments":[{"comment":"The title in the arXiv header reads 'TOW ARDS' with an extra space, and there are a few other spacing artifacts in the text; a copyedit pass would improve readability.","section":"Overall"},{"comment":"The choice of the nine nonlinear functions is inherited from [15], but no analysis or ablation is provided for the specific function set. A sentence explaining why this set is appropriate for the three camera parameters, or a small ablation, would strengthen the method section.","section":"Section 2.2"},{"comment":"The legend for 'Input feature from previous block' and 'Output feature from current block' is unclear because both arrows in the CPA-NAFBlock diagram use the same style, making the dataflow difficult to follow. Please use distinct arrow styles or labels.","section":"Fig. 2(a)"},{"comment":"The evaluation protocol for SID and SIDD should state explicitly how the patch-level PSNR/SSIM are computed (e.g., average over non-overlapping test patches) and how the full-frame results are produced; currently this information must be inferred from the note about Restormer.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of an image-processing venue and the idea is simple and relevant. The main risk is that the central claim is under-evidenced due to the parameter-count confound and the lack of prior-conditioning baselines. These are fixable with additional experiments, so I do not recommend rejection, but the revision must include capacity-matched and condition-ablated controls plus a clearer statement of what is being compared against."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it does something genuinely new for real denoising: it conditions a restoration network on ISO, shutter speed, and F-number through adaptive layer normalization, and shows you can then drive the denoiser's strength at inference. Second, the main quantitative claim is confounded by model capacity, and the authors don't do the experiment that would fix it.\n\nThe idea is simple and practical. The authors encode three camera parameters with nonlinear functions borrowed from ParamISP, feed that vector through adaLN into NAFNet blocks, and also show it works on Restormer. Code is released. That's a reasonable contribution for an ICIP paper, and the writing is honest: they acknowledge the parameter increase and note the SIDD gain is smaller in their discussion of ISO distributions.\n\nThe soft spot is the one the stress-test flagged. CPADNet has 9.45M parameters versus 6.67M for the baseline, a 42% increase, and there is no ablation that keeps the architecture and parameter count fixed while removing or scrambling the camera-parameter conditioning. So the +0.99 dB SID gain could come from the extra MLP capacity alone. On SIDD the gain is +0.02 dB, which is within single-run noise. The controllability figures only show the network responds to its input; they don't show the conditioning carries information beyond capacity. There are also no error bars or multiple seeds, and no comparison with prior noise-level conditioning baselines like CBDNet or DUDN, so \"improves state-of-the-art\" is doing more work than the experiments support.\n\nThat said, the paper is not broken. The central idea is plausible, the description of the noise model and exposure algorithm is sound, and the authors are careful to note the SID/SIDD discrepancy. The omission is a missing control, not a logical flaw.\n\nWho should read it: people working on conditional denoising or camera-aware restoration. It will get them thinking, but they should not take the SID gain at face value until a capacity-matched ablation exists. I would send this to peer review rather than desk reject: it's a borderline paper with a fixable weakness, and the referee can ask for the missing control. As is, it's a conditional accept at best.","headline":"A sensible but under-evidenced ICIP paper: camera-conditioned denoising via adaLN shows a real SID gain, but the authors never rule out that extra parameters do the work, and the SIDD gain is negligible.","tokens_in":8607,"tokens_out":1940,"would_cite":false,"duration_ms":23918,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Camera metadata can both improve and steer a real-image denoiser.","keywords":["controllable image denoising","camera parameters","adaptive layer normalization","real image denoising","ISO","shutter speed","F-number","EXIF metadata"],"falsifier":"Train a capacity-matched baseline (NAFBlock enlarged to 9.45M parameters, no camera conditioning) on the same SID patches; if its PSNR is not below CPADNet's 25.57 dB, the central claim that camera parameters drive the improvement is falsified.","tokens_in":7794,"feed_emoji":"📷","tokens_out":7660,"duration_ms":74918,"temperature":0.7,"pith_summary":"This paper sets out to show that the noise level of a real photograph is encoded well enough in three camera settings—ISO, shutter speed, and F-number—that a denoising network given those values as a condition outperforms the same network without them and can be adjusted by the user at inference time. The authors turn the three metadata values into a single conditioning vector and inject it into every block of a U-shaped denoising network through adaptive layer normalization. On the SID and SIDD real-image benchmarks the conditioned network, CPADNet, raises PSNR and SSIM over its baseline, and the same conditioning also improves a Transformer-based restorer. If the claim holds, metadata that already accompanies ordinary photo files becomes a free knob for choosing how much noise to remove.","feed_headline":"Camera metadata improves and tunes real-image denoising","feed_subtitle":"Feeding EXIF camera settings to a denoiser boosts quality and lets users dial denoising strength.","key_machinery":"The camera parameter vector $v \\in \\mathbb{R}^{27}$. For each of ISO, shutter speed, and F-number, nine non-linear functions—$x$, $1/x$, $\\sqrt{x}$, $x^{-1/2}$, $x^{1/4}$, $x^{-1/4}$, $\\log x$, $\\sin(\\log x)$, $\\cos(\\log x)$—are applied and normalized to $[0,1]$, then concatenated. This vector is fed to adaptive layer normalization (adaLN) layers, which predict channel-wise affine parameters $\\gamma$ and $\\beta$, replacing the plain layer normalization of NAFBlock with a conditioning mechanism; dropout at rate 0.2 is applied before the affine prediction. The vector encodes an approximate noise level and conditions every block of the U-shaped network, so the network's denoising strength responds to the supplied camera settings. For smartphone images lacking F-number, the F-number branch is replaced by a one-hot encoding of camera type passed through an embedding layer.","core_discovery":"The central claim is that camera parameters are valuable ground-truth information about real noise level, so conditioning a denoising network on them during training improves accuracy and grants interactive control at inference. Concretely, ISO, shutter speed, and F-number are mapped through nine non-linear functions, normalized to [0,1], and concatenated into a camera parameter vector $v \\in \\mathbb{R}^{27}$. The vector is consumed by adaptive layer normalization layers that produce per-channel scale and shift, replacing the plain layer normalization of the NAFBlock baseline in every block. At inference the user can replace the true metadata with a milder or harsher camera setting to lighten or strengthen denoising. The paper reports gains on SID (24.58 to 25.57 dB patch PSNR) and on SIDD (39.41 to 39.43 dB patch PSNR) over the baseline, and similar gains when the conditioning is added to Restormer.","pith_inferences":["An implication left implicit is that existing photo libraries could be re-rendered at different noise strengths by editing the metadata values fed to the network, since EXIF data already exists for most images.","A capacity-matched control—enlarging the baseline to CPADNet's 9.45M parameters without camera conditioning—would separate the contribution of the metadata from the contribution of extra parameters; the paper does not include such an experiment.","If the conditioning vector faithfully represents noise level, feeding an ISO between two training values should yield a denoised result whose strength lies between the two endpoints; the paper demonstrates the endpoints but not this interpolation curve."],"forward_implications":["A single trained CPADNet can produce lighter or stronger denoising at inference by substituting lower or higher ISO, shutter speed, or F-number values, with no retraining.","The conditioning also improves Restormer, a Transformer architecture, so it is not tied to the specific NAFBlock design.","Because the three parameters come from standard EXIF metadata, the method does not require extra capture-time instrumentation.","The improvement is larger on SID (about 1 dB) than on SIDD (about 0.02 dB), which the paper ties to the closer match between training and test ISO distributions on SID."],"supporting_citations":[{"why":"Supplies the Sony α7S II noisy/clean RAW pairs and the ISO, shutter speed, and F-number metadata used for training and full-image tests.","marker":"[2]"},{"why":"Supplies smartphone noisy/clean pairs and camera labels (ISO, shutter, camera type) used for training and validation/test evaluation.","marker":"[1]"},{"why":"Provides the NAFBlock restoration architecture whose normalization is replaced by adaptive layer normalization in CPADNet.","marker":"[7]"},{"why":"Supplies the adaptive layer normalization mechanism that injects the camera parameter vector as per-channel scale and shift.","marker":"[8]"},{"why":"Supplies the set of non-linear equalization functions used to map each camera parameter into a 9-dimensional vector.","marker":"[15]"},{"why":"Transformer denoising architecture used to show the camera-parameter conditioning transfers beyond the NAFBlock baseline.","marker":"[17]"},{"why":"Provides the heteroscedastic Gaussian noise model that connects read noise and shot noise to the choice of camera parameters.","marker":"[9]"}],"fun_headline_variants":["Camera settings dial in denoising strength and boost quality","ISO, shutter speed, and F-number tune image denoising","Use camera metadata to adjust denoising strength","Control real-image denoising with camera parameters","Dial denoising strength via camera settings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported gains come from the camera-parameter conditioning itself, not from the larger network: CPADNet has 9.45M parameters versus 6.67M for the baseline, and no capacity-matched ablation is included.","fun_headline_variants_meta":{"raw":{"variants":["Camera settings dial in denoising strength and boost quality","ISO, shutter speed, and F-number tune image denoising","Use camera metadata to adjust denoising strength","Control real-image denoising with camera parameters","Dial denoising strength via camera settings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00065,"raw_usage":{"total_tokens":2938,"prompt_tokens":858,"completion_tokens":2080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":2004}},"tokens_in":474,"tokens_out":2080,"duration_ms":18498,"temperature":1.0,"reasoning_tokens":2004,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:47:05.309123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a capacity-matched baseline (NAFBlock enlarged to 9.45M parameters, no camera conditioning) on the same SID patches; if its PSNR is not below CPADNet's 25.57 dB, the central claim that camera parameters drive the improvement is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies smartphone noisy/clean pairs and camera labels (ISO, shutter, camera type) used for training and validation/test evaluation."},{"cited_title":"Ffdnet: Toward a fast and flexible solution for cnn-based image denoising,","cited_arxiv_id":null,"evidence_quote":"Provides the NAFBlock restoration architecture whose normalization is replaced by adaptive layer normalization in CPADNet."},{"cited_title":"Toward convolutional blind denoising of real photographs,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive layer normalization mechanism that injects the camera parameter vector as per-channel scale and shift."},{"cited_title":"Uformer: A general u-shaped transformer for image restoration,","cited_arxiv_id":null,"evidence_quote":"Transformer denoising architecture used to show the camera-parameter conditioning transfers beyond the NAFBlock baseline."},{"cited_title":"Deep universal blind image denoising,","cited_arxiv_id":null,"evidence_quote":"Provides the heteroscedastic Gaussian noise model that connects read noise and shot noise to the choice of camera parameters."}],"review_version":1}