{"id":"d7362d58-1316-4ec9-900e-1fe897eabda0","arxiv_id":"2505.02705","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CRWKV, a bidirectional RWKV network with context-guided token shift and frequency mixing, matches or modestly exceeds prior real-image denoisers with lower computational cost.","lead":"The paper presents CRWKV, an RWKV-based image denoising model with a context-guided token shift and a frequency-mixing module. The authors report denoising quality on par with or slightly better than recent state-of-the-art methods on several real-world benchmarks, while using fewer FLOPs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim is contradicted by Table 3: CRWKV is slower than Restormer (62.74 ms vs 47.40 ms) while also slightly lower on SIDD PSNR, so the central practical-advantage claim is not established.","rationale":"The reader's weakest-assumption focuses on the fixed CTS offset dictionary from LG-BPN. That is a legitimate limitation, and the paper itself acknowledges that learnable offsets remain an open question in Section 4.3. I do not think it is the most load-bearing issue, however: the model already generalizes to three unseen datasets, and CTS contributes about 0.15-0.5 dB in ablations, so a domain mismatch would reduce but not necessarily destroy the contribution. The efficiency head-to-head is more decisive because the Abstract's central practical claim directly contradicts the paper's own Table 3. Since a reader going only by the Abstract would believe CRWKV is both more accurate and 40% faster, while the actual comparison against the strongest competitor (Restormer) is lower accuracy and 32% slower on SIDD, this is an overclaim that should be corrected. The reader's CONDITIONAL verdict already captures this risk, so no verdict change is proposed.","tokens_in":12811,"tokens_out":7395,"duration_ms":85050,"concrete_test":"Run an independent latency benchmark using the released CRWKV code and official Restormer and MambaIR checkpoints on the same RTX 4090, with identical input resolutions (e.g., 256x256, 512x512, and 1024x1024), batch size, warm-up, and torch.cuda.synchronize, averaging at least 20 runs. If CRWKV remains slower than Restormer at matched settings, then the abstract's 'inference time up to 40%' claim must be restricted to comparisons with SwinIR/ShuffleFormer, and the central SOTA-and-faster claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Abstract is that CRWKV outperforms state-of-the-art methods quantitatively and reduces inference time by up to 40%. For that claim to hold, CRWKV should be at least competitive in speed with the accuracy-relevant baselines. Table 3 of the paper reports CRWKV at 62.74 ms, Restormer at 47.40 ms, and MambaIR at 79.61 ms. Thus CRWKV is 32% slower than Restormer, not 40% faster. The 'up to 40%' reduction is only achieved relative to much weaker baselines such as ShuffleFormer (98.96 ms) or SwinIR (170.33 ms). On the primary SIDD benchmark, Table 2 also shows Restormer at 40.01 dB PSNR versus 39.87 dB for CRWKV. The paper's repeated wording 'outperforming SOTA methods' and 'reducing inference time up to 40%' is therefore materially overbroad unless explicitly scoped to the less accurate baselines and to non-SIDD datasets. The method may still be a useful contribution, but the advertised practical advantage is the weakest load-bearing part of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CRWKV, a U-shaped RWKV-based architecture for real-world image denoising. The contribution is a combination of three mechanisms: BiWKV, a bidirectional linear-complexity token-mixing operation adapted from Vision-RWKV; CTS, a fixed context-guided token shift whose offset dictionary is taken from an external Pearson-correlation analysis; and FMix, an FFT-based frequency filtering module. The model is trained on SIDD and evaluated on SIDD, ccnoise, PolyU, and a synthetic Urban100GP benchmark, with ablations over module configuration, shift mechanisms, layer depth, and loss functions. The manuscript claims that CRWKV outperforms state-of-the-art methods quantitatively and reduces inference time by up to 40%.","tokens_in":13084,"tokens_out":3676,"duration_ms":42895,"significance":"If the empirical results hold, CRWKV would be a useful linear-complexity denoiser with competitive accuracy and a smaller parameter/FLOP footprint than Restormer and MambaIR on several real-noise benchmarks. The paper's strengths include a public code release, multiple unseen-dataset evaluations, and fairly extensive ablations. The central practical-advantage claim, however, is not fully supported by the reported numbers: CRWKV does not beat Restormer on SIDD PSNR and is not faster than Restormer in the reported inference-time comparison. The fixed CTS offset dictionary also raises a transferability question, though the authors explicitly acknowledge this limitation. Overall the architecture is plausible and the contribution is meaningful, but the advertised claims need to be scoped and the efficiency evidence needs more detail before publication.","major_comments":[{"comment":"The abstract's claim that CRWKV \"outperforms the state-of-the-art methods quantitatively and reduces inference time up to 40%\" is overbroad. On the primary SIDD benchmark, Table 2 shows Restormer at 40.01 dB PSNR versus CRWKV at 39.87 dB, and Table 3 shows CRWKV at 62.74 ms inference time versus Restormer at 47.40 ms. Thus CRWKV is neither the most accurate on SIDD nor faster than Restormer. The claim should be explicitly scoped to the datasets on which CRWKV leads (ccnoise, PolyU, Urban100GP) and to speed comparisons against baselines such as ShuffleFormer and SwinIR, from which the up-to-40% reduction appears to be derived.","section":"Abstract; Table 2; Table 3"},{"comment":"The computational-complexity analysis is not sufficient to support the inference-efficiency claim. Table 3 reports only a single millisecond value per method with no stated input resolution, GPU thermal state, warm-up procedure, or number of repeated runs; without this information the measured ordering (CRWKV slower than Restormer) cannot be attributed to architecture rather than implementation or protocol. The authors should report the measurement setup and ideally provide speed curves over input sizes, which would also make Figure 1(b)'s memory comparison interpretable relative to the time claim.","section":"Section 4.2; Table 3; Figure 1(b)"},{"comment":"The generality of the CTS mechanism is not tested against the variability of real-world noise statistics. The fixed offset dictionary in Algorithm 1 is taken from the Pearson-correlation analysis in LG-BPN, and the paper does not provide evidence that this specific noise-neighborhood geometry transfers to SIDD, ccnoise, PolyU, or Urban100GP beyond the aggregate scores. A concrete test would be to vary or learn the offset dictionary on each noise regime, or to report CTS performance under synthetic noise with different correlation structures; the current ablation in Table 5 only compares two fixed dictionaries derived from the same external prior.","section":"Section 3.2; Section 4.3; Table 5"},{"comment":"Algorithm 1 is not fully reproducible as written. The channel allocation o[c:c+k*w_p] uses the real number k*w_p, where k=C/psum and w_p=1/d_p, but Python slicing requires integer bounds; the paper does not specify how fractional channel counts are rounded or how the final channel segment is handled when C is not exactly divisible by the sum of weights. Since CTS is a central component, the authors should give the exact integer allocation rule or refer to the public code for this detail.","section":"Algorithm 1; Section 3.2"}],"minor_comments":[{"comment":"The sentence \"The proposed FRWKV method was evaluated\" contains a typo: the model is called CRWKV, not FRWKV.","section":"Section 4.2"},{"comment":"The row labels in Table 4 are ambiguous because the checkmarks are not mapped to the configurations described in Section 4.3. Please label each row explicitly, e.g., \"FMix only,\" \"FMix + CTS(CRM),\" \"FMix + CTS(CMix),\" and \"full CRB,\" so the ablation deltas can be verified.","section":"Table 4"},{"comment":"The visual example in Figure 1(a) is labeled Urban100, while the quantitative benchmark in Table 2 is Urban100GP with different PSNR/SSIM values for the same methods; the relationship between these two versions should be clarified.","section":"Figure 1(a) vs. Table 2"},{"comment":"The number of neighboring pixels used in the base CTS configuration is never stated; the paper only says the CTS(+) variant covers 16 neighbors at Manhattan distance 3. Please give the base shift set and its Manhattan distances.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"No additional concerns beyond those in the report. The paper fits the journal's scope, and the issues are addressable with revised claims and additional experimental detail."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead the CRWKV paper. The actual new work is modest but real: combining RWKV with a fixed context-guided token shift borrowed from LG-BPN's noise correlation analysis and a frequency-mix (FMix) branch for high-frequency attenuation. The BiWKV variant with symmetric position bias is a sensible adaptation for images. The paper is a solid engineering study, not a conceptual breakthrough. The ablations are thorough, the code is public, and the experiments span four datasets. That is credit where it's due.\n\nThe soft spots are concentrated in the claims. The abstract says \"outperforming the existing state-of-the-art methods quantitatively and reducing inference time up to 40%.\" Table 2 shows Restormer at 40.01 dB SIDD vs. 39.87 for CRWKV, so the quantitative claim holds only if you ignore SIDD PSNR. Table 3 shows CRWKV at 62.74 ms vs. Restormer's 47.40 ms, so it is 32% slower, not 40% faster; the 40% reduction is relative to much weaker baselines. The authors do acknowledge the SIDD PSNR exception in Section 4.2, which helps, but the abstract and repeated wording still overreach. That needs fixing.\n\nA second, quieter concern is the fixed offset dictionary from LG-BPN. The paper assumes the noise-neighborhood geometry measured in that external study transfers to SIDD, ccnoise, PolyU, and Urban100GP. It may, but the paper does not test robustness to different noise correlation structures. The authors mention avoiding learnable offsets to reduce overfitting, which is a reasonable choice, but they leave the transfer question open.\n\nMinor issues: Section 4.2 calls the model \"FRWKV\" once, and the efficiency table doesn't state the input resolution used for timing, which makes the numbers hard to interpret.\n\nOn balance, the paper deserves a serious referee. The architecture is plausible, the experiments are extensive, and the code availability helps reproducibility. My recommendation: send to peer review, but require the authors to scope the abstract claims to what the tables actually show and to add a discussion of when the CTS dictionary might fail.\n\nBest,\n[You]","headline":"A solid RWKV-based denoising architecture paper whose abstract overstates both accuracy and speed; worth reviewing after the claims are scoped to the data.","tokens_in":13645,"tokens_out":1947,"would_cite":true,"duration_ms":21048,"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":"The paper proposes CRWKV, a linear-complexity denoiser that matches transformer-level quality on real-world noise while using fewer parameters and less memory.","keywords":["image denoising","real-world noise","RWKV","linear complexity","frequency-domain feature","token shift","multi-view learning"],"falsifier":"Take a noisy dataset whose noise correlation is deliberately different from the CTS offset dictionary—for example, noise with long-range or anisotropic spatial correlation—and compare CRWKV against a variant that learns the offset dictionary from that dataset. If the fixed-offset model loses substantially while the learnable variant does not, the transfer assumption behind CTS is false; the same test can be done by measuring the Pearson correlation neighborhood on ccnoise or PolyU and checking whether it matches the offsets used.","tokens_in":12598,"feed_emoji":"🖼️","tokens_out":9148,"duration_ms":92774,"temperature":0.7,"pith_summary":"CRWKV is a new architecture for real-world image denoising built on the RWKV sequence model rather than on transformers or state-space scans. The paper claims that three mechanisms—a context-guided token shift that follows the measured correlation structure of real noise, a frequency-mix branch that separates high-frequency noise from detail, and a bidirectional WKV operator that lets every pixel interact with every other pixel in linear time—let a 20.19M-parameter model match or outperform leading denoisers on ccnoise, PolyU, and Urban100GP, and nearly tie the best transformer on SIDD while using less memory and, the paper claims, reducing inference time by up to 40% relative to several state-of-the-art baselines. This matters because the strongest existing denoisers are transformer-based and computationally heavy; a linear-complexity alternative with comparable quality would make high-resolution and resource-constrained denoising more practical. The model is trained on SIDD with an L1 loss and is shown to generalize to unseen real-noise datasets without retraining.","feed_headline":"CRWKV denoiser beats state-of-the-art on real-world noise","feed_subtitle":"A 20M-parameter linear-complexity model rivals transformer denoisers on real noise","key_machinery":"The load-bearing mechanism is the Context Receptance Block, which stacks Frequency Mix (FMix) and Context Receptance Module (CRM) branches around Channel Mix layering. Context-guided Token Shift (CTS) allocates channels to shifted versions of neighboring pixels using a fixed offset dictionary drawn from prior Pearson-correlation analysis of real noise, weighting each offset by the reciprocal of its Manhattan distance; this fixes the receptive-field shape to the noise's correlation structure. The Frequency Mix module applies a 2D FFT, weights the frequency response linearly, passes it through Leaky ReLU, and modulates the original feature map with the inverse-FFT result, so noise concentrated in high frequencies is selectively attenuated. The Bidirectional WKV (BiWKV) operator, with relative position bias $b_{t,i}=-(|t-i|-1)/T$ and a learnable current-token bonus $u$, computes a weighted sum over all pixels in linear time while treating forward and backward neighbors symmetrically, overcoming the causal restriction of standard WKV.","core_discovery":"The central claim is that CRWKV, a U-shaped encoder-decoder with long-skip connections, outperforms every compared state-of-the-art method on ccnoise, PolyU, and Urban100GP, and reaches 39.87 dB PSNR on SIDD, just 0.14 dB below Restormer's 40.01 dB, while using fewer parameters and one of the smallest FLOP counts among modern denoisers. The paper attributes this to the interaction of its three designed modules: CTS encodes the spatial correlation geometry of real noise directly into token mixing, FMix attenuates noise-dominated frequency bands, and BiWKV provides symmetric, linear-complexity global pixel mixing that avoids the asymmetry of causal scans. Power-spectrum analysis of feature maps shows CRWKV retains high-frequency amplitude more steadily across encoder layers than MambaIR, which the paper takes as evidence that the model preserves detail while suppressing noise throughout the network.","pith_inferences":["If CTS's fixed offset dictionary is the source of its gains, then re-estimating that dictionary per target domain (or per camera) could improve generality further; the paper's own CTS(+) ablation shows the offset set has a sweet spot, so adaptive offsets are a natural next test.","The BiWKV operator's spatial symmetry is a design that likely transfers to other low-level vision tasks such as super-resolution and deblurring, though the paper only evaluates denoising.","The paper's power-spectrum comparison suggests that stable amplitude retention across layers is a useful diagnostic for restoration backbones; extending that analysis to test-time noise level or camera type could make the claim more precise.","The reported inference-time savings should be checked at high resolutions and on edge hardware, because the paper's memory and FLOP curves imply the advantage grows with input size, making mobile deployment a promising but untested extension."],"forward_implications":["A linear-complexity sequence backbone can match transformer-based restoration quality, so high-resolution images can be denoised with moderate GPU memory and FLOPs rather than quadratic attention cost.","The fixed, correlation-derived shift dictionary shows that noise-specific local geometry can be hard-coded into a denoiser without learned offsets, simplifying training and reducing overfitting risk.","Frequency-domain mixing at deep, semantically rich layers preserves high-frequency detail better than at middle layers, giving a practical design rule for similar restoration networks.","A single SIDD-trained CRWKV transfers to other real-noise domains such as ccnoise, PolyU, and synthetic Urban100GP, indicating that the learned noise model is not dataset-specific.","The L1 loss, combined with these modules, outperforms MSE, Charbonnier, and PSNR losses on the tested benchmarks, so fidelity-focused training is preferred for real noise."],"supporting_citations":[{"why":"Supplies the Pearson-correlation analysis of noise neighborhoods from which CTS's fixed offset dictionary is derived.","marker":"[Wang et al., 2023]"},{"why":"Provides the RWKV architecture and WKV computation that CRWKV extends with the bidirectional operator.","marker":"[Peng et al., 2023]"},{"why":"Vision-RWKV adaptation that introduces quad-shift and Bi-WKV concepts which CRWKV builds on.","marker":"[Duan et al., 2024]"},{"why":"Restormer is the main transformer baseline whose SIDD PSNR and efficiency the paper compares against.","marker":"[Zamir et al., 2022]"},{"why":"MambaIR is the state-space baseline used for feature-map power-spectrum analysis and efficiency comparison.","marker":"[Guo et al., 2025]"},{"why":"SIDD supplies the training data and the primary real-noise test set.","marker":"[Abdelhamed et al., 2018]"},{"why":"AP-BSN represents the self-supervised blind-spot baseline that CRWKV is compared with.","marker":"[Lee et al., 2022]"},{"why":"BM3D is the classical non-learning denoising baseline in the comparison table.","marker":"[Dabov et al., 2007]"}],"fun_headline_variants":["CRWKV denoiser: linear-complexity, 40% faster, beats SOTA on real noise","Real-noise denoising: CRWKV beats transformers with 40% less inference time","CRWKV: compact multi-view model that beats heavy denoisers on real noise","20M-parameter CRWKV: linear attention outperforms transformer denoisers","CRWKV: fast, accurate real-noise denoising with multi-view learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that the spatial pattern of noise correlation measured in one prior study is the same across all real-world noise, because CTS fixes that pattern in advance rather than learning it per dataset.","fun_headline_variants_meta":{"raw":{"variants":["CRWKV denoiser: linear-complexity, 40% faster, beats SOTA on real noise","Real-noise denoising: CRWKV beats transformers with 40% less inference time","CRWKV: compact multi-view model that beats heavy denoisers on real noise","20M-parameter CRWKV: linear attention outperforms transformer denoisers","CRWKV: fast, accurate real-noise denoising with multi-view learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3021,"prompt_tokens":939,"completion_tokens":2082,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1966}},"tokens_in":555,"tokens_out":2082,"duration_ms":18909,"temperature":1.0,"reasoning_tokens":1966,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:42:44.944387+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a noisy dataset whose noise correlation is deliberately different from the CTS offset dictionary—for example, noise with long-range or anisotropic spatial correlation—and compare CRWKV against a variant that learns the offset dictionary from that dataset. If the fixed-offset model loses substantially while the learnable variant does not, the transfer assumption behind CTS is false; the same test can be done by measuring the Pearson correlation neighborhood on ccnoise or PolyU and checking whether it matches the offsets used.","supporting_citations":[{"cited_title":"Lg-bpn: Local and global blind-patch network for self-supervised real-world denoising","cited_arxiv_id":null,"evidence_quote":"Supplies the Pearson-correlation analysis of noise neighborhoods from which CTS's fixed offset dictionary is derived."},{"cited_title":"Mambair: A simple baseline for image restoration with state-space model","cited_arxiv_id":null,"evidence_quote":"MambaIR is the state-space baseline used for feature-map power-spectrum analysis and efficiency comparison."},{"cited_title":"A high-quality denoising dataset for smartphone cameras","cited_arxiv_id":null,"evidence_quote":"SIDD supplies the training data and the primary real-noise test set."},{"cited_title":"Ap-bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network","cited_arxiv_id":null,"evidence_quote":"AP-BSN represents the self-supervised blind-spot baseline that CRWKV is compared with."},{"cited_title":"Image denoising by sparse 3-d transform-domain collaborative filtering","cited_arxiv_id":null,"evidence_quote":"BM3D is the classical non-learning denoising baseline in the comparison table."}],"review_version":1}