{"id":"40a673f5-27e2-4478-8d73-6765224b49ce","arxiv_id":"2412.20157","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-granularity mixture-of-experts image restoration model that routes each degraded image to an expert using both degradation and granularity estimates, outperforming all-in-one baselines.","lead":"UniRestorer groups image degradations into clusters at several zoom levels, trains a separate restoration expert for each cluster, and chooses which expert to use based on both the predicted degradation and how uncertain that prediction is. It reports large gains over other all-in-one restoration models on seven tasks and mixed corruptions, narrowing the gap to task-specific models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing premise that Eq. (4) K-means clusters are degradation-semantic is never quantitatively established; a random-cluster control and cluster-purity check are needed before the central mechanism can be credited.","rationale":"The reader's weakest assumption names exactly the premise I consider load-bearing: the Eq. (4) clusters must be semantically coherent for the proposed degradation-specific experts and granularity routing to be the mechanism behind the reported gains. I agree with that assessment. I found no internal mathematical contradiction in the paper; the concern is an unverified empirical premise. The paper does have useful independent support: the extractor ablation (Tab. 6), the granularity ablations (Tabs. 7-8), the routing loss ablations (Tab. D-E), and the scaled-baseline control (Tab. B) all suggest the design choices matter. However, none of these tests establishes that the fine clusters align with degradation type or severity; a random partition with matched capacity could in principle yield similar improvements. Because code and trained models are not released, this premise cannot currently be checked externally. The placeholder 'Frobnication' references [5-8,35-36] are a manuscript-cleanliness issue but not load-bearing. I therefore recommend keeping the reader's CONDITIONAL verdict; the condition should include a quantitative demonstration of cluster semantics, e.g., the random-cluster control and cluster-purity check described above.","tokens_in":23157,"tokens_out":8414,"duration_ms":93292,"concrete_test":"Train a matched control MoE with the same expert architectures, cluster sizes, and router training, but assign training samples randomly to the same 19/7 hierarchical clusters instead of using Eq. (4). Compare average PSNR on Table 1 and Table 2; if the random-cluster control is within about 0.3 dB, the semantic content of the clusters is not driving the gains. Additionally, compute adjusted Rand index (or purity) of the actual Eq. (4) clusters against the known synthetic severity bins (rain strength 0-50/50-100/100-150, sigma 0-15/15-35/35-50, haze beta, and blur kernel size) from Sec. A; low agreement would corroborate that the clusters are arbitrary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (4) hierarchical K-means on the DA-CLIP DR space is the foundation of the claimed mechanism: fine-grained experts are supposed to specialize on semantically coherent degradations, and granularity estimation is supposed to route to a coarser expert when degradation estimation is unreliable. If the 19 fine clusters are not aligned with actual degradation type/severity, the experts are just arbitrary capacity partitions, and the reported gains could come from the 27 full-size expert networks plus sparse routing rather than from degradation-aware specialization. The paper provides only qualitative t-SNE (Fig. B) and a deraining-only extractor ablation (Tab. 6); it never reports cluster purity, agreement with the known synthetic degradation parameters of Sec. A, or a random-cluster control. Because Hd and Hg in Eq. (7) are trained against the same cluster centers, they cannot independently validate the semantics of the partition. This is the load-bearing empirical premise of the abstract's claim, and it is currently unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniRestorer, an all-in-one image restoration framework built around a multi-granularity mixture-of-experts design. A DA-CLIP-based degradation extractor is trained on fine-grained synthetic degradations; its features are hierarchically K-means clustered into three granularity levels (e.g., 1/7/19 clusters for single-degradation and 1/4/8 for mixed-degradation). Separate full-network experts are trained on each cluster's data, and routing is performed by two learned heads: a degradation head that selects a finest-level cluster and a granularity head that chooses an appropriate granularity level, intended to be robust to degradation estimation errors. Experiments cover seven single-degradation tasks and seven mixed-degradation scenarios, plus real-world and unseen-degradation generalization. The reported results substantially outperform prior all-in-one methods in auto mode and approach or exceed single-task models in an instruction mode where the task name is provided.","tokens_in":23347,"tokens_out":7024,"duration_ms":71895,"significance":"If the main empirical claims hold, UniRestorer is a significant advance for all-in-one restoration: the auto-mode gains in Table 1 (e.g., Rain100L 41.68 dB vs. 38.57 dB and SOTS 36.44 dB vs. 31.34 dB for the best prior methods) and in Table 2 are large, and the real-world and unseen-degradation results in Tables 4-5 are encouraging. The paper also includes useful ablations (Tables 6-8), a router-component study (Table E), scaled-up baseline comparisons (Table B), and lightweight expert variants (Table A), which help separate the contribution of the proposed mechanism from raw capacity. The authors commit to releasing code and models. The central weakness is that the semantic validity of the learned degradation clusters—the load-bearing premise of the whole mechanism—is not quantitatively established, and the comparison to single-task models in instruction mode is not fully fair as presented.","major_comments":[{"comment":"The load-bearing premise that K-means clusters in the DA-CLIP DR space are semantically coherent—corresponding to degradation types and severities—is never quantitatively established. The paper provides only qualitative t-SNE (Fig. B) and a deraining-only DR-extractor ablation (Table 6); neither measures agreement between cluster assignments and the known synthetic degradation labels (type and degree ranges in Sec. A/B). Because the degradation and granularity heads Hd/Hg and the routers Gd/Gg are trained against these same cluster centers in Eqs. (6)-(7), the routing results cannot independently validate the partition. I request cluster purity or adjusted mutual information against the known degradation parameters, plus a control in which experts are trained on random partitions of the same cardinality. Without this, the reported gains could come from the large set of full-size experts and sparse routing rather than from degradation-aware specialization.","section":"Sec. 3.2, Eq. (4), and Sec. B"},{"comment":"The robustness claim—that granularity estimation routes to coarser experts when degradation estimation is unreliable—is not directly tested. egran is trained from the distance between edeg and the finest-level center, but no experiment verifies that egran tracks actual degradation estimation error or that coarser routing is selected in high-error cases. Table C reports expert usage statistics and routing accuracy (92%/82%), but not error-conditional behavior. Please add a controlled analysis, e.g., degrade inputs with known parameters, measure edeg error against ground truth, and show that Gg selects coarser experts as error increases; or perturb edeg and show that UniRestorer degrades less than a finest-only router. This is needed to support the abstract's central claim of robustness to degradation estimation error.","section":"Sec. 3.3, Eq. (7), and Table C"},{"comment":"The instruction-mode comparison with single-task models is not on equal footing. For deraining, the authors retrain experts on Rain200H/Rain200L/DID/DDN, which is a larger and more diverse set than the Rain100L used in the all-in-one comparison, while the single-task baselines are trained on their own standard data. Also, Ours† receives the task name as a pruning signal during inference. To support the 'closing the gap' claim, either train the single-task baselines on the same enlarged deraining data, or report Ours† without the extra deraining data, and state clearly in the main text what training data each entry uses.","section":"Sec. 4.2, Table 3, and Sec. A.1"}],"minor_comments":[{"comment":"Equation (2) appears to add Gaussian noise after the Softmax; the standard noisy top-k gating from Shazeer et al. applies the noise to the logits before the Softmax. Please correct or clarify, since the current equation does not implement the cited mechanism.","section":"Sec. 3.1, Eq. (2)"},{"comment":"The abbreviation 'MiO' is not defined in the main text, and the ablation tables report small differences (e.g., 24.46 vs. 24.41 in-dist for 3 vs. 4 granularity levels) without error bars or multiple seeds, so the claimed plateau and optimal-configuration conclusions are not statistically supported.","section":"Tables 7-8"},{"comment":"The reference list contains placeholder entries ([5]-[8], [35], [36]) and missing citations such as 'DesnowNet [? ]' in Sec. 1 and Sec. 2. Please clean up the bibliography and all in-text citations before publication.","section":"References and citations"},{"comment":"The PromptIR row is cited as [82], which is TransWeather; the PromptIR reference is [67]. Please check the citation and also define 'CAR' and 'MiO' at first use.","section":"Table 2"},{"comment":"The notation in Eq. (6) is under-specified: the candidate set {F0,0,...,Fn−1,k} and the TopK operation need precise definitions, and the sentence 'Gd is conducted in the finest-grained level' is inconsistent with Gg selecting the granularity. Please clarify the two-stage routing procedure.","section":"Sec. 3.3, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper is empirically strong and the proposed mechanism is plausible, but the requested cluster-validity and error-robustness analyses are necessary before the central claims can be fully credited. The single-task comparison also needs to be made fair or clearly qualified. The presence of placeholder references is a production-quality concern that should be fixed. If the authors can supply the requested controls and clarifications, the paper could be suitable for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a real step forward for all-in-one restoration. The multi-granularity MoE plus granularity-estimation router is genuinely new relative to single-granularity explicit-MoE methods like GRIDS and RestoreAgent, and the reported gains are large: on Rain100L it goes from ~38.6 dB for the best prior all-in-one to 41.7 dB, and on SOTS from ~31.3 to 36.4 dB. Those numbers are enough to take the architecture seriously.\n\nWhat's good: the experiments are unusually broad — seven single-degradation tasks, seven mixed-degradation settings with in-dist and out-dist parameters, plus real-world and unseen-corruption generalization. The ablations in Tables 6–8 support the design choices (DR extractor choice, number of clusters, number of granularity levels). The instruction-mode comparison against single-task models is fair and informative. The supplementary overhead analysis shows sparse inference is cheap even though training has 339M parameters; the scaled-up Restormer-Large and PromptIR-Large baselines in Table B are a good-faith attempt to separate capacity from routing quality.\n\nSoft spots. First, the central mechanism is underevidenced. The paper claims fine-grained experts specialize on semantically coherent degradations because the DR space is K-means clustered at multiple granularities. But the paper never shows that the resulting clusters agree with the known synthetic degradation parameters (type and severity) used to generate the training data. The t-SNE is qualitative, and the extractor ablation is on deraining only. The routing accuracy numbers (92%/82%) are computed against the same cluster centers, so they cannot validate the partition independently. A random-cluster control and a cluster-purity measure against the known parameter ranges would directly address this. Without it, one alternative explanation is that the gains come largely from 27 full-size experts plus sparse routing, not from degradation semantics.\n\nSecond, there are no error bars anywhere, and several hyperparameters (the degree thresholds for the DR extractor labels) are hand-set. Third, the abstract and intro promise code and models but nothing is released yet. Fourth, the manuscript is not fully cleaned: placeholder 'Frobnication' references, duplicated citations, and at least one dangling citation mark. Minor, but shouldn't be there.\n\nThe soft spots are real but not fatal. The core claim holds up as an empirical advance, and the ablations do show that multi-granularity helps beyond simply adding experts. The load-bearing premise is testable, and the paper would be much stronger with that test.\n\nWho it's for: anyone working on all-in-one restoration, MoE architectures, or degradation-aware low-level vision. I'd bring it to reading group and would cite it once the code is out. Recommendation: accept for peer review, with a request for the cluster-semantics control, error bars, and code release before final acceptance.","headline":"UniRestorer is a serious empirical advance in all-in-one restoration, but the multi-granularity cluster semantics need direct validation before the mechanism is fully credited.","tokens_in":23879,"tokens_out":3579,"would_cite":true,"duration_ms":32689,"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":"UniRestorer claims an all-in-one restorer can get degradation-specific precision and robustness to estimation error at once by routing at a granularity matched to estimation confidence.","keywords":["all-in-one image restoration","mixture of experts","degradation representation","granularity estimation","hierarchical clustering","expert routing","image restoration generalization"],"falsifier":"Train the same multi-granularity MoE pipeline after replacing the learned degradation embeddings with random features or with features from an encoder never exposed to degradations; if the gains persist, the claimed benefit does not come from degradation structure. A complementary test is to evaluate routing accuracy on degradation parameters drawn from outside the hand-set strength bins used to train the extractor — the paper reports 92% routing accuracy in-distribution and 82% out-of-distribution, so a large drop under shifted parameters would show the granularity estimator only interpolates the training bins.","tokens_in":22960,"feed_emoji":"🧩","tokens_out":8515,"duration_ms":78927,"temperature":0.7,"pith_summary":"This paper tries to establish that an all-in-one image restoration system can have both strengths at once: the precision of degradation-specific models and the resilience of degradation-agnostic ones. Its solution is to represent image degradation at several granularities, train separate restoration experts for each cluster in that hierarchy, and add a granularity estimator that measures how much the degradation estimate can be trusted. Confident estimates are routed to fine-grained specialists; uncertain ones are routed to coarser, more general experts. If the claims hold, universal restoration would largely close the performance gap to dedicated single-task models across rain, haze, noise, blur, low light, snow, and compression artifacts, while also generalizing to mixed and unseen corruptions.","feed_headline":"Expert routing closes all-in-one restoration's single-task gap","feed_subtitle":"Confident degradation estimates route to specialists; uncertain ones to generalists — across seven restoration tasks.","key_machinery":"The load-bearing object is the multi-granularity degradation set: a hierarchy of non-overlapping clusters obtained by K-means on degradation embeddings extracted by a DA-CLIP-trained encoder, with the number of groups growing from 1 to 7 to 19 in the single-degradation setup, and from 1 to 4 to 8 in the mixed-degradation setup. Each group trains its own full-network expert, so the same corrupted image can be restored by a specialist or by a generalist depending on routing. The router that carries the argument is the pair of estimators $H_d$ and $H_g$ trained with the data-uncertainty loss $L_{\\mathrm{dg}} = \\frac{1}{2 e_{\\mathrm{gran}}}(u_{y_i} - e_{\\mathrm{deg}})^2 + \\frac{1}{2}\\ln e_{\\mathrm{gran}}$, where $e_{\\mathrm{gran}}$ plays the role of a learned error estimate for $e_{\\mathrm{deg}}$; a second router then picks the coarsest expert whose breadth matches that uncertainty. A load-balance loss prevents the routing from collapsing onto the finest-grained experts. The mechanism turns a single point estimate of degradation into a confidence-aware routing decision.","core_discovery":"The central claim is that degradation estimation error need not be a fatal flaw of degradation-aware restoration: it can be modeled and exploited. UniRestorer first trains a fine-grained degradation-representation extractor based on DA-CLIP, then runs hierarchical K-means on the extracted embeddings to build a three-level degradation set — 1, 7, and 19 groups in the single-degradation setup, 1, 4, and 8 in mixed-degradation — and trains a full restoration network for every group. Two routers then decide the expert: one estimates the finest-level degradation group, and the other estimates a granularity that, through a data-uncertainty loss, encodes how wrong the first estimate is likely to be. In the authors' experiments the automatic mode reaches 41.68 dB PSNR on Rain100L deraining and 36.44 dB on SOTS dehazing versus 38.57 and 31.34 dB for the strongest compared all-in-one method, and the instruction mode matches or beats dedicated single-task models on most tasks. The paper itself notes that its training data come from public datasets and its degradation space is synthetic; it lists larger-scale data and real-world degradation alignment as future work.","pith_inferences":["Beyond the paper, the confidence-conditioned routing recipe is transferable: any model that conditions on a noisy predictor could use a similar uncertainty estimate to fall back to a broader hypothesis class, such as blind super-resolution or restoration agents managing a zoo of specialist models.","Because the authors freeze experts before training routers and show gains over scaled-up monolithic models, an economical extension would replace full experts with LoRA or adapter specialists and test whether granularity routing still delivers most of the gain at a fraction of the training cost.","A natural next step the paper leaves implicit is an online variant that updates cluster centers as new degradation types arrive, turning the fixed multi-granularity partition into a growing taxonomy of corruptions."],"forward_implications":["On the seven single-degradation tasks, automatic routing lifts the five-task average PSNR to 33.38 dB from 30.58 dB for the best compared all-in-one method, and the seven-task average to 31.34 dB; expert specialization, not a bigger shared backbone, drives the gain.","Instruction mode, where the user supplies the degradation type as a pruning mask, puts UniRestorer at parity with or above single-task models on most tasks, showing the all-in-one versus single-task performance gap is largely a routing problem rather than a capacity problem.","Mixed-degradation results improve from 22.06 to 24.46 dB in-distribution and from 17.23 to 19.45 dB out-of-distribution over the shared backbone, and the multi-granularity ablation shows coarse levels are what buy out-of-distribution robustness.","The system generalizes to real-world and unseen datasets (LHP, LOLv2, RealSnow, raindrop, under-display camera, underwater) without retraining, consistent with the claim that coarse experts provide a fallback when the degradation is outside the training distribution."],"supporting_citations":[{"why":"Supplies the contrastive degradation-representation extractor backbone whose embeddings are clustered into the multi-granularity degradation set.","marker":"[59]"},{"why":"Supplies the K-means clustering procedure used to build the hierarchical non-overlapping degradation groups.","marker":"[62]"},{"why":"Provides the base transformer restoration architecture used as the expert network for most degradation groups.","marker":"[100]"},{"why":"Precedent for dividing a complete degradation space into subspaces and training full-model experts, which the paper builds on and compares against.","marker":"[12]"},{"why":"Precedent for using full restoration networks as experts and for the evaluation setting used in all-in-one comparisons.","marker":"[14]"},{"why":"Supplies the data-uncertainty learning formulation used to train the granularity-estimation branch and the L_dg loss.","marker":"[13]"},{"why":"Supplies the noisy top-k routing formulation and the load-balance loss used to train the routers.","marker":"[75]"}],"fun_headline_variants":["Adaptive granularity routing turns degradation error into a feature","Two routers pick the right restoration expert, closing the single-task gap","All-in-one restoration that knows when to trust its degradation estimate","UniRestorer: error-aware expert selection beats all-in-one and nears single-task","Granularity-aware experts: all-in-one restoration closes single-task gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline rests on the premise that the degradation extractor's feature clusters correspond to degradation types and severities that are genuinely useful for restoration; if the clusters are arbitrary, the finer experts receive inconsistent training data and the routers can only learn an arbitrary mapping.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive granularity routing turns degradation error into a feature","Two routers pick the right restoration expert, closing the single-task gap","All-in-one restoration that knows when to trust its degradation estimate","UniRestorer: error-aware expert selection beats all-in-one and nears single-task","Granularity-aware experts: all-in-one restoration closes single-task gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000882,"raw_usage":{"total_tokens":3835,"prompt_tokens":991,"completion_tokens":2844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2764}},"tokens_in":607,"tokens_out":2844,"duration_ms":22963,"temperature":1.0,"reasoning_tokens":2764,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:29:00.707778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same multi-granularity MoE pipeline after replacing the learned degradation embeddings with random features or with features from an encoder never exposed to degradations; if the gains persist, the claimed benefit does not come from degradation structure. A complementary test is to evaluate routing accuracy on degradation parameters drawn from outside the hand-set strength bins used to train the extractor — the paper reports 92% routing accuracy in-distribution and 82% out-of-distribution, so a large drop under shifted parameters would show the granularity estimator only interpolates the training bins.","supporting_citations":[{"cited_title":"Controlling vision-language models for universal image restoration","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive degradation-representation extractor backbone whose embeddings are clustered into the multi-granularity degradation set."},{"cited_title":"Machine learning, volume 1","cited_arxiv_id":null,"evidence_quote":"Supplies the K-means clustering procedure used to build the hierarchical non-overlapping degradation groups."},{"cited_title":"Outrageously large neural networks: The sparsely-gated mixture-of-experts layer","cited_arxiv_id":null,"evidence_quote":"Supplies the noisy top-k routing formulation and the load-balance loss used to train the routers."}],"review_version":1}