{"id":"9711286b-31a3-4826-a1b7-771bb1dcad93","arxiv_id":"2412.10651","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LAN learns a pixel-wise correction to a noisy image so a frozen denoiser sees noise closer to its training distribution, improving cross-dataset denoising.","lead":"A frozen image-denoising network can be adapted to unfamiliar noise by optimizing a per-pixel offset added to the input, instead of retraining the network. The LAN method gives modest but consistent PSNR gains on two real-world datasets, though its core mechanism is only qualitatively validated.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.2's core assumption that minimizing Eq. 12 pulls input noise toward pretraining distribution Ds is unverified; reported gains may arise from a different mechanism, so the claimed noise-distribution-matching mechanism is not established.","rationale":"The reader's weakest_assumption identifies exactly the same gap: the paper assumes, without proof or quantitative evidence, that minimizing Eq. 12 drives the adapted input's noise distribution toward Ds. This is the central load-bearing element because it justifies the method's design (adapting the input instead of the network) and explains why LAN should outperform network-level adaptation. My concern focuses on this same unresolved point and proposes a concrete, feasible test that would settle it. Since the reader has already assigned a CONDITIONAL verdict based on this and other issues, and my independent review does not identify a more severe flaw that would require rejection, the verdict should remain CONDITIONAL (UNCHANGED). The empirical results are consistent and include multiple backbones and datasets, but the small gains without error bars and the untested mechanism warrant the conditional status. The proposed test is directly executable with the paper's released code and synthetic noise protocol, making it a practical verification step.","tokens_in":12942,"tokens_out":7278,"duration_ms":69529,"concrete_test":"In the synthetic setup of Section 4.5 (known Ds and Du, ground-truth clean image xu available), run LAN on a test set of images with noise from Du. For each image, compute the empirical noise before adaptation (n0 = yu − xu) and after adaptation (n1 = yu + φ* − xu). Estimate the Wasserstein-1 distance (or KL divergence via histograms) between n0 and samples from Ds, and between n1 and samples from Ds. Compare the distances across a held-out set of ≥50 images. If the distance does not significantly decrease for a majority of images while PSNR improves, the central mechanism is falsified. Additionally, repeat on real datasets (SIDD→PolyU/Nam) using a learned noise model for Ds (e.g., a normalizing flow trained on SIDD noise residuals) to obtain a distributional distance, and report the correlation between distance reduction and PSNR improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that minimizing Eq. 12 with respect to the pixel-wise offset φ brings the input noise closer to the training noise distribution Ds, thereby improving the frozen denoiser's performance on unseen noise. This assumption (stated in Section 3.2) is load-bearing because the entire novelty and interpretation of LAN depend on it: the method adapts the input rather than the network precisely so that the noise becomes more like Ds. Yet the paper provides no proof, and the only evidence is a qualitative histogram (Figure 6) on synthetic Gaussian/Gamma noise. The argument in Section 4.2 that 'noise adaptation is not just additional denoising' is based on a visual/PSNR comparison of the adapted noisy image, not on a quantitative measure of how close the adapted noise actually is to Ds.\n\nConcretely, the self-supervised loss is a consistency loss between two downsampled views, not a direct measure of distributional similarity to Ds. There is no guarantee that its minimizer corresponds to −ϵ from the decomposition eu = es + ϵ (Eq. 5); many other φ could reduce the loss—for example, by smoothing the image, suppressing high-frequency content, or exploiting the downsampling operators D1/D2. The offset φ is unconstrained and pixel-wise, so it can encode structural content that is not 'noise-like'. If the mechanism is not actually distribution matching, the empirical gains might be explained by input smoothing or by the network being evaluated on a slightly altered image, and the method might not generalize to noise types where such trivial manipulation is ineffective. Thus the unsupported assumption directly affects the validity of the proposed approach and the interpretation of all reported improvements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LAN, a test-time adaptation method for image denoising that keeps a pretrained denoiser frozen and instead optimizes a pixel-wise additive offset on the given noisy image using a self-supervised loss such as ZS-N2N or Neighbor2Neighbor. The adapted image is then passed through the frozen denoiser. Experiments compare LAN with full-network, first-layer, last-layer, and meta-learning adaptation on SIDD-pretrained DnCNN, Restormer, and Uformer evaluated on PolyU and Nam, reporting consistent PSNR/SSIM gains and reduced runtime/memory relative to full fine-tuning. A qualitative histogram experiment on synthetic Gaussian and Gamma noise is offered as evidence that the offset moves the input noise toward the training noise distribution.","tokens_in":13269,"tokens_out":6020,"duration_ms":53953,"significance":"If the empirical gains are robust, LAN is an interesting and orthogonal contribution: it avoids modifying network weights, is memory-efficient for large backbones, and opens an input-side view of test-time adaptation. The paper ships code and evaluates three backbones, two self-supervised losses, and two target real-world datasets, which is a strength. The evaluation is not circular, since the offset is optimized with a self-supervised loss while performance is measured on clean PSNR/SSIM. However, the significance is limited by the unverified mechanism and the lack of statistical controls; without these, the contribution is an empirical technique whose working principle remains unsupported.","major_comments":[{"comment":"The central claim that minimizing the self-supervised loss with respect to the offset pulls the input noise toward the training distribution Ds is asserted but not demonstrated. Equation (12) is a consistency loss between two downsampled views after the frozen denoiser, not a distributional divergence between noise residuals. An unconstrained pixel-wise offset can lower this loss by smoothing the image or by exploiting the downsampling operators D1/D2, without making the residual statistically similar to the training noise. The only supporting evidence is Figure 6, which shows marginal histograms for one synthetic setup, and Figure 4, which is qualitative. Matching marginal histograms does not establish spatial noise statistics. Please provide a quantitative distributional measure (e.g., MMD or KL divergence on estimated residuals) before and after adaptation, and preferably an ablation where the offset is constrained or where a different loss is optimized, to show that the mechanism is noise-distribution matching rather than generic input alteration.","section":"3.2, Eq. (12)"},{"comment":"The central empirical claim of consistent improvement lacks statistical support. No error bars, standard deviations, or significance tests are reported, and the comparisons use per-method learning rates chosen by the authors (Section 4.1), so the relative gains could depend on hyperparameter tuning. Please report variance across images or runs, a sensitivity analysis over learning rates and iteration counts, and ideally paired significance tests. Without these, differences of 0.1-0.3 dB may not be distinguishable from run-to-run variability, and the claim that LAN 'consistently outperforms' full-trainable adaptation is not fully supported.","section":"Table 1"},{"comment":"The argument that full-trainable adaptation is fundamentally weaker than LAN is based on a single longer-iteration curve (Uformer with ZS-N2N on Nam). Since the comparison is sensitive to optimization schedules, this one instance is not enough to establish that full-trainable adaptation cannot reach comparable performance with more iterations or better tuning. Please show similar curves for the other backbone/loss/target combinations or explain why this instance is representative. Otherwise the conclusion remains limited to the tuned 20-iteration protocol.","section":"Section 4.2, Figure 5"}],"minor_comments":[{"comment":"The initialization of the offset phi is not specified; if it is initialized to zero, state so explicitly, since the optimization is otherwise not fully specified and the semantics of the learned offset depend on the starting point.","section":"Section 4.1"},{"comment":"The caption states that 'Full-trainable and LAN (Ours) finetuned the pretrained network via ZS-N2N', which is inaccurate for LAN, since LAN does not modify the network parameters; please rephrase to clarify what is adapted in each method.","section":"Figure 3 caption"},{"comment":"References [17] and [18] are the same Neighbor2Neighbor paper with the same authors and venue; please merge them and renumber the citations accordingly.","section":"References"},{"comment":"The header 'Time Memory' is ambiguous; please label the columns as runtime ratio and memory ratio with explicit units or a note that they are percentages relative to the full-trainable baseline.","section":"Table 2"},{"comment":"The zero-shot comparison is not controlled for training budget or initialization: a randomly initialized DnCNN is trained for more than 1K iterations per image, whereas the other methods use a SIDD-pretrained network adapted for at most 20 iterations. This limits the strength of the conclusion that the large training set is the decisive factor; please add a matched-budget comparison or soften the claim.","section":"Section 4.3, Table 3"}],"recommendation":"major_revision","confidential_remarks":"This is a promising but not yet fully supported submission. The empirical results are suggestive, and the experimental breadth is a strength, but the claimed mechanism needs quantitative validation and the headline comparisons need error bars. I would not reject: the direction is novel and the paper ships code. The main risk is that the gains arise from input smoothing or from exploiting the self-supervised loss rather than from genuine noise-distribution matching, which the authors can address with residual-distribution statistics and appropriate ablations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper does something new. Instead of fine-tuning the denoiser to an unseen noise distribution, it freezes the network and optimizes a per-pixel additive offset on the noisy input, using the same self-supervised loss (ZS-N2N or Neighbor2Neighbor) that others use for test-time adaptation. The empirical claim holds up across three backbones (DnCNN, Restormer, Uformer), two self-supervised losses, and two unseen-noise datasets, with consistent gains of roughly 0.2–0.8 dB PSNR over the pretrained baseline. For a subfield that mostly tweaks architectures or training losses, this is a distinct and useful direction.\n\nWhat the paper does well: the idea is clearly motivated and explained; the evaluation is reasonably thorough, including comparisons to full-network, first-layer, last-layer, and meta-learning adaptation; and the authors include a small sanity check (Section 4.5, Figure 6) showing that adapted noise looks closer to the training noise distribution. The code is public. The paper is also honest about its main limitation: the number of trainable parameters scales with input image size.\n\nWhere the soft spots are: the load-bearing assumption in Section 3.2—that minimizing the self-supervised loss pulls the input noise toward the pretraining distribution D_s—is not actually established. The evidence is one qualitative histogram on synthetic Gaussian/Gamma noise and visual inspection. The loss is a consistency between downsampled views; its minimizer does not obviously correspond to removing the deviation e_s→u. It could be smoothing or other content-dependent manipulation. This does not kill the empirical result, but it means the mechanism story is speculative. The comparison also lacks error bars or variance estimates, and the per-method learning rates are hand-tuned, which matters when gains are under 1 dB. The meta-learning baseline is described in less detail than the others, so the comparison may not be perfectly fair.\n\nI want to push back on one worry in the stress-test note: the circularity concern is actually mild. Performance is measured on clean PSNR/SSIM, not on the self-supervised objective, so the evaluation is not directly fitting the metric.\n\nBottom line: this is a solid, well-scoped contribution for researchers working on test-time adaptation in image restoration. It deserves a serious referee. The authors should be asked to either demonstrate the distribution-matching mechanism with a quantitative distance measure on real noise, or soften the claim and present LAN purely as a practical TTA method. Either way, it deserves review time.","headline":"A simple, genuinely new test-time adaptation trick—freeze the denoiser and learn a per-pixel input offset—with consistent but modest gains, and a mechanism story that is plausible but not yet proven.","tokens_in":13811,"tokens_out":2209,"would_cite":true,"duration_ms":21758,"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":"Adding a learnable pixel-wise offset to a noisy input, while keeping the pretrained denoiser frozen, improves denoising on unseen noise.","keywords":["image denoising","test-time adaptation","self-supervised learning","input noise adaptation","unseen noise","real-world noise","learnable offset","blind denoising"],"falsifier":"Compute a statistical distance such as KL divergence or maximum mean discrepancy between the residual noise of the LAN-adapted image and the noise distribution the denoiser was trained on, across the PolyU and Nam test sets; if PSNR improves while this distance does not shrink, the mechanism is unsupported. A second check is to run LAN on images whose noise already matches the training distribution; if it still gives the same gain, the improvement is not explained by noise-distribution matching.","tokens_in":12746,"feed_emoji":"🖼️","tokens_out":4876,"duration_ms":37922,"temperature":0.7,"pith_summary":"This paper claims that when a pretrained denoiser meets an image with unseen noise, the better move is to adapt the input rather than the network: add a learnable pixel-wise offset to the noisy image so its noise moves closer to the distribution the denoiser was trained on. The authors keep the denoiser frozen and optimize only the offset with a self-supervised loss, and report consistent PSNR and SSIM gains over full-network, first-layer, last-layer, and meta-learning adaptation across DnCNN, Restormer, and Uformer on real-world noise datasets. If true, this reframes test-time adaptation as input-side correction, which is cheaper and avoids the overfitting that degrades fine-tuned networks, and it opens a research direction orthogonal to self-supervised denoising.","feed_headline":"Adapt the noise, not the network, to denoise unseen images","feed_subtitle":"A learnable input offset pushes unseen noise toward what a pretrained denoiser expects, beating full network fine-tuning.","key_machinery":"The central object is the pixel-wise noise offset $\\phi$ added directly to the noisy input. It is trained by minimizing a self-supervised denoising loss, specifically the Zero-Shot Noise2Noise and Neighbor2Neighbor losses, with the pretrained denoiser held frozen. The offset is meant to approximate $-\\epsilon_{s\\to u}$, the deviation that separates the unseen noise from the noise the network was trained on, thereby translating the input into $y_{u\\to s} = y_u + \\phi \\approx x_u + e_s$. This input-side translation carries the whole argument: all adaptation happens in the image, not in the network weights.","core_discovery":"The central claim is that unseen noise $e_u$ can be written as seen noise $e_s$ plus a deviation $\\epsilon_{s\\to u}$, and that subtracting a learned approximation of that deviation from the input image, rather than changing the pretrained denoiser, lets the frozen network denoise as if the noise were familiar. Formally, the method optimizes $\\phi = \\arg\\min_\\phi \\| f_{\\theta^*}(D_1(y_u+\\phi)) - D_2(y_u+\\phi) \\|_2^2$ with $f_{\\theta^*}$ frozen, then estimates the clean image as $f_{\\theta^*}(y_u+\\phi^*)$. The paper reports that this outperforms adapting all or part of the network, with the largest gain on Restormer going from 38.03 dB to 38.86 dB on SIDD-to-Nam with the ZS-N2N loss, and shows qualitatively that the adapted image's noise histogram shifts toward the training noise distribution.","pith_inferences":["A natural extension is to apply input-side adaptation to other restoration tasks such as super-resolution, deblurring, or low-light enhancement, where a pretrained model faces a distribution shift in the degradation.","The per-pixel offset could be constrained to a low-rank or smooth parametric form to remove the image-size-dependent memory cost the authors list as a limitation.","Combining LAN with a small amount of model adaptation may yield gains larger than either alone, since the offset fixes input statistics while fine-tuning could correct remaining network-side bias.","One could test whether the learned offset behaves like an anti-adversarial perturbation by measuring its norm and spatial structure and comparing it to adversarial noise of similar magnitude."],"forward_implications":["A pretrained denoiser can be adapted to a new noise type without any weight updates, removing the risk of overwriting or overfitting the network during test-time adaptation.","The input-side offset produces gains within 5 to 20 adaptation iterations across all tested backbones and both self-supervised losses, with the largest reported gain on Restormer from 38.03 dB to 38.86 dB on SIDD-to-Nam with ZS-N2N.","LAN is computationally cheaper than full-network adaptation for 256x256 images, using roughly 74 to 93 percent of the runtime and 74 to 93 percent of the memory of full-trainable adaptation in the reported settings.","Because the offset is optimized per image with a frozen network, the method is orthogonal to the choice of self-supervised loss and could be stacked on top of future self-supervised objectives."],"supporting_citations":[{"why":"Provides the SIDD dataset, which supplies the training noise distribution for the pretrained denoisers.","marker":"[1]"},{"why":"Supplies the Zero-Shot Noise2Noise self-supervised loss used to train the learnable offset.","marker":"[35]"},{"why":"Supplies the Neighbor2Neighbor self-supervised loss used as an alternative training objective.","marker":"[17]"},{"why":"Provides the Restormer backbone, a pretrained denoiser evaluated in the experiments.","marker":"[53]"},{"why":"Provides the Uformer backbone, a pretrained denoiser evaluated in the experiments.","marker":"[46]"},{"why":"Provides the DnCNN backbone, retrained on SIDD for the experiments.","marker":"[54]"},{"why":"Provides the Nam dataset used as an unseen-noise test set.","marker":"[37]"},{"why":"Provides the PolyU dataset used as an unseen-noise test set.","marker":"[48]"},{"why":"Provides the MAML meta-learning baseline that LAN is compared against in adaptation.","marker":"[14]"}],"fun_headline_variants":["Frozen net, adapted noise: LAN wins on unseen images","Add learnable noise offset, keep denoiser frozen","Tweak the input, not the model, to denoise unseen noise","LAN adapts noise, leaving the network untouched"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that minimizing the self-supervised loss on the adapted input actually pulls the noise distribution toward the training distribution; the paper asserts this in Section 3.2 and supports it only with one qualitative histogram, without a proof or quantitative measure.","fun_headline_variants_meta":{"raw":{"variants":["Frozen net, adapted noise: LAN wins on unseen images","Add learnable noise offset, keep denoiser frozen","Tweak the input, not the model, to denoise unseen noise","LAN adapts noise, leaving the network untouched"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000488,"raw_usage":{"total_tokens":2420,"prompt_tokens":975,"completion_tokens":1445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1376}},"tokens_in":591,"tokens_out":1445,"duration_ms":10172,"temperature":1.0,"reasoning_tokens":1376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:44:46.445375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute a statistical distance such as KL divergence or maximum mean discrepancy between the residual noise of the LAN-adapted image and the noise distribution the denoiser was trained on, across the PolyU and Nam test sets; if PSNR improves while this distance does not shrink, the mechanism is unsupported. A second check is to run LAN on images whose noise already matches the training distribution; if it still gives the same gain, the improvement is not explained by noise-distribution matching.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SIDD dataset, which supplies the training noise distribution for the pretrained denoisers."},{"cited_title":"Zero-shot noise2noise: Efficient image denoising without any data","cited_arxiv_id":null,"evidence_quote":"Supplies the Zero-Shot Noise2Noise self-supervised loss used to train the learnable offset."},{"cited_title":"Neighbor2neighbor: Self-supervised de- noising from single noisy images","cited_arxiv_id":null,"evidence_quote":"Supplies the Neighbor2Neighbor self-supervised loss used as an alternative training objective."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the Restormer backbone, a pretrained denoiser evaluated in the experiments."},{"cited_title":"Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising","cited_arxiv_id":null,"evidence_quote":"Provides the DnCNN backbone, retrained on SIDD for the experiments."},{"cited_title":"A holistic approach to cross-channel im- age noise modeling and its application to image denoising","cited_arxiv_id":null,"evidence_quote":"Provides the Nam dataset used as an unseen-noise test set."},{"cited_title":"Model- agnostic meta-learning for fast adaptation of deep networks","cited_arxiv_id":null,"evidence_quote":"Provides the MAML meta-learning baseline that LAN is compared against in adaptation."}],"review_version":1}