{"id":"df94c644-b3a1-4597-ad0e-bdf1ce0658a3","arxiv_id":"2512.10948","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A cluster-guided mixture-of-experts network with frequency modulation reports competitive all-in-one image restoration results, with uneven gains and no public code.","lead":"ClusIR is a new deep-learning system that restores photos damaged by noise, haze, rain, blur, or low light using one shared model. It groups images into learned 'degradation clusters' and then routes each image to specialized experts and frequency-based refinements, reporting small gains over prior all-in-one restorers on several benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PCGRM's causal role is contradicted by the paper's own One-by-One results: Sec. 4.4 states cluster prototypes fail to activate, yet Table 3 reports SOTA single-task numbers, so the central mechanism may not drive the reported gains.","rationale":"The reader already identified the Sec. 4.4 contradiction and the unverified causal role of PCGRM as the weakest assumption, and returned a CONDITIONAL verdict. My stress-test narrows this to a sharper internal inconsistency: the paper's own text says the cluster prototypes fail to activate in One-by-One training, yet Table 3 reports SOTA results in that setting. That is not merely a theoretical worry about unimodality or over-parameterization; it is direct textual evidence that the central mechanism is not required for the reported performance in at least one major experimental setting. A matched flat-router ablation and cluster-activation diagnostics would settle whether PCGRM contributes beyond its parameters. This does not change the verdict: the empirical claims remain plausible but conditional on resolving this contradiction and releasing code/error bars. I do not see grounds to reject outright, because the All-in-One tables and ablations are internally consistent and the mechanism could still matter in joint training; nor do I see grounds to accept, because the central novelty is currently unsupported by causal evidence. Hence UNCHANGED (conditional).","tokens_in":19985,"tokens_out":3184,"duration_ms":35342,"concrete_test":"Run a controlled replacement: keep WTB+DAFMM and the same expert count/training budget, but replace PCGRM (Eqs. 3-10) with a standard flat softmax MoE router (Eqs. 1-2) of matched capacity, under identical One-by-One and N+H+R protocols. If the One-by-One results on Kodak24/SOTS/Rain100L/LOL remain within ~0.1 dB of Table 3 while the All-in-One gains shrink, then PCGRM is not the driver of the claimed single-task SOTA. Complement this by measuring cluster-activation statistics (e.g., entropy of p(C=c|x), top-1 probability, per-cluster expert usage) on the released One-by-One checkpoints: if top-1 probabilities are near-uniform or expert selection is identical across clusters, the Sec. 4.4 failure is confirmed and PCGRM's causal benefit is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ClusIR's PCGRM—the hierarchical two-stage cluster-guided router—disentangles degradation recognition from expert activation and thereby delivers SOTA all-in-one restoration. The paper's Sec. 4.4 explicitly says that under One-by-One training ClusIR \"fails to effectively activate the cluster prototypes, which undermines the PCGRM, prevents reliable degradation discrimination, and consequently limits the generation of semantic prompts.\" Yet Table 3 reports that the same ClusIR achieves the best or near-best AiOIR results on Kodak24 denoising (35.06 dB at sigma=15), SOTS dehazing (32.67/0.983), and LOL low-light (23.82/0.852). This is an internal inconsistency: if the load-bearing mechanism is disabled or ineffective, the SOTA numbers in that setting cannot be attributed to it. Either the One-by-One results are produced by the other components (WTB, DAFMM, or simply model capacity), or the mechanism is not actually inert despite the text. In both cases, the causal contribution of PCGRM—the paper's main novelty—is unverified. The All-in-One ablation (Table 4) shows +0.27 dB when PCGRM-MoE is added, but this only demonstrates correlation with the full pipeline, not that the two-stage probabilistic routing, rather than the extra parameters/experts, causes the gain. Since Contribution (1) and the abstract tie SOTA performance to the cluster-guided design, this contradiction is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ClusIR proposes a cluster-guided all-in-one image restoration framework. Its two main components are a Probabilistic Cluster-Guided Routing Mechanism (PCGRM), which factorizes expert routing into a degradation-cluster posterior p(C=c|x) and a cluster-conditional expert posterior p(E=e|c,x), and a Degradation-Aware Frequency Modulation Module (DAFMM), which uses cluster prompts with DWT and a learnable Frequency Self-Mining Block for frequency decomposition. The paper reports experiments on three-degradation (N+H+R) and five-degradation (N+H+R+B+L) settings, on the composite CDD11 benchmark in the supplementary material, and on One-by-One single-task training. Table 1 reports an average PSNR/SSIM of 33.06/0.923 for the N+H+R setting, claimed to surpass recent all-in-one methods; Table 2 reports 30.58/0.919 on five tasks. The central claim is that PCGRM's hierarchical two-stage routing disentangles degradation recognition from expert activation and thereby delivers state-of-the-art adaptive restoration.","tokens_in":20474,"tokens_out":5822,"duration_ms":55882,"significance":"If the mechanism's causal role were established, ClusIR would offer a useful design pattern for AiOIR: explicit learnable cluster prototypes with two-stage routing, coupled with frequency-domain modulation. The empirical strengths are the external benchmark evaluations (Tables 1–3 and supplementary CDD11 results), the component ablations, the orthogonal-initialization study, and the efficiency variants (ClusIR-T/S/Base). The gains over strong recent methods such as MoCE-IR, DFPIR, and AdaIR are modest but consistent, especially in average PSNR/SSIM. However, the load-bearing attribution of these gains to PCGRM is not yet supported: the paper's own Sec. 4.4 states that cluster prototypes fail to activate in One-by-One training, and the main ablation in Table 4 adds PCGRM-MoE as a whole package without isolating the two-stage routing from added parameters. The result is promising but needs substantial verification before the 'state-of-the-art via cluster-guided routing' claim is accepted.","major_comments":[{"comment":"The paper's own Sec. 4.4 states that in One-by-One training ClusIR 'fails to effectively activate the cluster prototypes, which undermines the PCGRM, prevents reliable degradation discrimination, and consequently limits the generation of semantic prompts.' Table 3 nonetheless reports best AiOIR results for Kodak24 (35.06 dB at sigma=15), SOTS (32.67/0.983), and LOL (23.82/0.852). This is an internal inconsistency that is load-bearing: Contribution (1)/(2) attributes state-of-the-art performance to the cluster-guided design, but the single-task SOTA cannot be attributed to PCGRM if the mechanism is inert. Please add controlled results under One-by-One with PCGRM ablated (WTB+DAFMM only) and with a matched-capacity single-stage MoE, and quantify cluster activation (e.g., entropy or purity against task labels) rather than relying only on t-SNE.","section":"§4.4 and Table 3"},{"comment":"The central motivation that a flat softmax router is 'unimodal' and therefore collapses mixed degradations is not established. A softmax over experts is a categorical distribution over the expert set; for different inputs x it can place high mass on different experts, so the router as a function of x can represent multimodal degradation mixtures. The limitation is a capacity/expressivity matter, not a consequence of unimodality. Please provide a formal argument or a controlled experiment comparing the two-stage PCGRM with a single-stage softmax router of matched parameters; otherwise the stated advantage of the factorization is a rhetorical contrast rather than a verified property.","section":"§3.2, Eqs. (1)–(2)"},{"comment":"The ablation from (a) to (b) adds PCGRM-MoE, which includes all expert parameters and gating weights, yielding +0.27 dB. This does not isolate the two-stage probabilistic routing from the increase in model capacity. A single-stage MoE baseline (Eq. (1)) with the same number of experts and parameters would disentangle capacity from the routing structure. Without such a baseline, the result is consistent with the hypothesis that the gain comes from extra parameters rather than from cluster-guided disentanglement.","section":"Table 4 (ablation)"},{"comment":"The claim that clusters encode degradation semantics rests on t-SNE and affinity maps generated by the same model. These visualizations show separation but not that the clusters correspond to the degradation classes used in training/testing, which would be necessary for 'explicit degradation semantics.' Please report quantitative alignment measures (e.g., cluster purity or NMI with task labels, or confusion between noise levels and tasks) and, ideally, a probe of p(C|x) against held-out degradation labels.","section":"§4.3, Figs. 5–6"}],"minor_comments":[{"comment":"The row labeled 'InstructIR [28] ECCV'24' cites reference [28] (Up-Restorer); InstructIR is reference [8]. Please correct the citation.","section":"Table 2"},{"comment":"Eq. (7) sums the Gaussian prompt over all N prototypes, while Eqs. (5)–(6) normalize over the top-K1 selected clusters. Please make the summation range consistent (selected set vs. full bank) and define sigma_c and epsilon_c precisely.","section":"§3.2, Eq. (7)"},{"comment":"The supplementary says denoising is evaluated on CBSD68 and Urban100, while Table 3 of the main text reports Kodak24. Please state which benchmark is used for One-by-One denoising and include Urban100 results or remove the mention.","section":"§6.1 / Table 3"},{"comment":"The figure captions 'Visual degradation of PCGRM-MoE' likely mean 'Visualization of PCGRM-MoE activations/affinity'; please reword.","section":"Figs. 6 and 9"},{"comment":"Eq. (14)'s IFFT(,) notation and Eq. (15)'s channel-wise gating are not fully defined; specify dimensions and the activation function used.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope and contains useful empirical results. The Sec. 4.4 contradiction and the capacity-control issue are fixable with additional experiments and rephrasing; I recommend major revision rather than rejection. If the authors cannot provide the matched-capacity single-stage baseline, the novelty claim should be weakened accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: ClusIR is a competent MoE-style all-in-one restoration paper with two new modules—PCGRM (hierarchical cluster-guided routing) and DAFMM (frequency self-mining + modulation)—and it reports competitive numbers on N+H+R, five-task, and CDD11 composited degradations. The gains over strong baselines are real but incremental: 0.2–0.4 dB on average, with a more noticeable +0.37 dB over WGWSNet on CDD11. The engineering is reasonable, ablations are present, and the supplementary includes small/tiny variants and initialization comparisons. It deserves a serious referee, but it has a load-bearing problem in its own text.\n\nThe main issue: Sec. 4.4 admits that under One-by-One training, ClusIR 'fails to effectively activate the cluster prototypes' and that this 'undermines the PCGRM, prevents reliable degradation discrimination, and consequently limits the generation of semantic prompts.' Yet Table 3 shows ClusIR beating prior all-in-one methods on exactly those single-task benchmarks. Either the numbers are not driven by PCGRM, or the mechanism is not actually inert. Either way, the paper's central contribution—that cluster-guided routing is what delivers the gains—is not supported in that setting. The ablation (Table 4) shows +0.27 dB from adding PCGRM-MoE, but that only correlates with the full pipeline, not with the two-stage factorization specifically.\n\nThere are also smaller issues. The motivation claims a flat softmax router is 'unimodal,' which is mathematically sloppy—a softmax can assign high probability to multiple experts. The paper doesn't release code or report error bars, which is standard for the field but matters when gains are small. The 'state-of-the-art across multiple heterogeneous benchmarks' in the abstract overstates what Table 2 shows (second-best PSNR, best SSIM).\n\nWhat's actually good: the CDD11 results on mixed degradations are the most interesting, and the t-SNE/affinity visualizations in the All-in-One setting do show cluster separation. The DAFMM with learnable frequency decomposition is a sensible extension of existing frequency-modulation ideas. The paper is well-written and positions itself clearly against MoE-based restoration work.\n\nVerdict: this is a useful paper for people working on all-in-one restoration, but the causal claim about PCGRM needs to be re-examined. A revision that either explains why single-task results don't need PCGRM, or reports ablations that isolate the two-stage factorization (e.g., compare against a single-level MoE with the same parameter count), would resolve the contradiction. Send it to review; a competent referee will push on exactly this point.","headline":"Solid all-in-one restoration paper with real but incremental gains; the central claim about PCGRM is undermined by the paper's own One-by-One results.","tokens_in":20925,"tokens_out":4934,"would_cite":false,"duration_ms":41495,"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":"ClusIR claims that making an all-in-one restorer explicitly recognize degradation clusters before choosing experts yields state-of-the-art unified restoration, with 33.06 dB average PSNR on noise, haze, and rain.","keywords":["All-in-one image restoration","mixture of experts","degradation-aware routing","cluster prototypes","frequency modulation","image denoising","image dehazing","image deraining"],"falsifier":"Train the identical architecture but replace PCGRM with a single flat softmax router over the same experts, holding parameter count roughly equal, and compare on CDD11 mixed degradations; if the flat router matches or beats ClusIR, the two-stage factorization is not doing the work. Alternatively, feed images with known single degradations and inspect the cluster posterior: if one cluster dominates regardless of degradation type, the clusters are not encoding degradation semantics.","tokens_in":19897,"feed_emoji":"🖼️","tokens_out":4684,"duration_ms":47050,"temperature":0.7,"pith_summary":"ClusIR argues that all-in-one image restoration improves when the network first identifies which kind of degradation it faces, then chooses specialized experts conditioned on that identification. The paper's route is a two-stage probabilistic router: a softmax over learned degradation clusters followed by a cluster-conditional softmax over experts, paired with a frequency module that uses cluster-derived cues to adjust low-frequency structure and high-frequency texture separately. On the three-task benchmark (denoising, dehazing, deraining) the system reports 33.06 dB average PSNR and 0.923 SSIM, ahead of the compared unified methods, and on the five-task setting it matches the best PSNR while leading SSIM. A sympathetic reader would take the contribution as evidence that explicit degradation semantics, not just implicit expert competition, is a useful organizing principle for unified restoration.","feed_headline":"Cluster-guided router lifts unified image restoration to 33.06 dB","feed_subtitle":"Two-stage cluster-then-expert routing separates degradation types before restoration, improving mixed-degradation recovery without extra sup","key_machinery":"Probabilistic Cluster-Guided Routing Mechanism (PCGRM): a two-stage router that first computes a softmax posterior over learned degradation clusters (spherically normalized, nearly orthogonal prototypes per encoder stage), then computes a cluster-conditional softmax over experts inside the top-K clusters, so expert choice is conditioned on a recognized degradation type. The Degradation-Aware Frequency Modulation Module (DAFMM) takes the cluster-derived prompt and applies a learnable low-pass filter (Frequency Self-Mining Block) plus Fourier amplitude-phase fusion to refine low frequencies and channel-gated modulation for high frequencies. Together they carry the claim that semantic degradati","core_discovery":"The paper claims that factorizing the router into two stages — first a degradation-aware cluster posterior p(C=c|x), then a cluster-conditional expert posterior p(E=e|c,x) — separates degradation recognition from expert activation. Flat single-softmax routers, it argues, force a unimodal decision space that entangles mixed degradations and creates unstable global expert competition. The two-stage form yields a multimodal distribution over experts and, combined with cluster-guided frequency modulation, gives the best reported average performance among compared all-in-one methods on the three-task benchmark (33.06 dB / 0.923) and strong composite-degradation results on CDD11 (27.33 dB / 0.878)","pith_inferences":["If the clusters genuinely correspond to degradation types, the mechanism doubles as a degradation classifier; one could use the cluster posterior to report \"this image looks hazy\" with no extra head. The paper does not exploit this, but it follows directly from the design.","The two-stage factorization is an instance of a general cost-geometry idea: first partition the input space coarsely, then act within the partition. Similar cluster-conditioned routing could apply to other inverse problems such as super-resolution or compression-artifact removal, where mixed corruptions are common.","The one-by-one result the paper reports — strong performance even when clusters do not activate — suggests the gains on single degradations may come substantially from added capacity, and the cluster mechanism's distinctive value is probably confined to mixed and composite degradations.","A controlled test that equalizes parameter count between the two-stage router and a single softmax router over the same expert set would isolate whether the factorization itself, rather than extra parameters, explains the benchmark gaps."],"forward_implications":["Unified restoration networks no longer need to trade away per-degradation specialization: the same weights can route differently for noise, haze, and rain while sharing most of the backbone.","Cluster assignments give a built-in, testable description of what degradation the model thinks it is seeing, which can be read out at inference time.","Because routing is conditioned on clusters, adding a new degradation type or benchmark may require only a new prototype rather than a retrained router.","Frequency modulation tied to cluster cues lets the same model emphasize structural low frequencies for blur-like degradations and texture high frequencies for noise-like ones."],"fun_headline_variants":["Cluster-guided router separates degradations, hits 33.06 dB","Two-stage routing separates degradation types for better restoration","Cluster-guided frequency modulation adapts to mixed degradations","ClusIR: separating degradation recognition from expert activation","Cluster-guided all-in-one restoration hits 33.06 dB average"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that two-stage cluster-then-expert routing genuinely disentangles degradation recognition from expert activation rather than re-parameterizing a single softmax; the paper's own Sec. 4.4 concedes that in one-by-one training the cluster prototypes fail to activate yet results remain strong, so the causal role of the clusters is not yet established.","fun_headline_variants_meta":{"raw":{"variants":["Cluster-guided router separates degradations, hits 33.06 dB","Two-stage routing separates degradation types for better restoration","Cluster-guided frequency modulation adapts to mixed degradations","ClusIR: separating degradation recognition from expert activation","Cluster-guided all-in-one restoration hits 33.06 dB average"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000581,"raw_usage":{"total_tokens":2568,"prompt_tokens":734,"completion_tokens":1834,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":1753}},"tokens_in":478,"tokens_out":1834,"duration_ms":12329,"temperature":1.0,"reasoning_tokens":1753,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:57:42.951395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical architecture but replace PCGRM with a single flat softmax router over the same experts, holding parameter count roughly equal, and compare on CDD11 mixed degradations; if the flat router matches or beats ClusIR, the two-stage factorization is not doing the work. Alternatively, feed images with known single degradations and inspect the cluster posterior: if one cluster dominates regardless of degradation type, the clusters are not encoding degradation semantics.","supporting_citations":[],"review_version":1}