{"id":"e37f6ec9-f044-4371-8ea1-48308401bfda","arxiv_id":"2411.18466","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MoCE-IR uses mixture-of-experts layers where experts have different complexity and a complexity-biased router assigns tasks to the right expert, improving all-in-one image restoration.","lead":"This paper introduces an image restoration model that routes each degraded photo to one of several specialty 'experts' that differ in compute and receptive field. A bias toward simpler experts makes the model automatically send easy tasks like deraining to cheap experts and hard tasks like dehazing to expensive ones, improving both speed and accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The expert scaling formula in Sec. 3.1 contradicts the text: r=C/2^i makes expert 1 the largest, while the bias b assumes expert 1 is the lightest, so the claimed 'complexity' ordering is ambiguous.","rationale":"The reader correctly identified the routing bias as the least secure premise, focusing on parameter count versus actual computational cost. Our stress test sharpens this into a concrete internal inconsistency: as printed, Eq. 3.1 and the surrounding text disagree about which expert is lightest, and consequently about the direction of the bias in Eq. 6. This is load-bearing because the entire contribution is the claim that a complexity bias produces task-discriminative experts; the ablation in Tab. 5b confirms the bias is necessary, yet the paper does not provide a consistent scalar definition of expert complexity. The issue is not fatal in principle—the reported numbers could be reproducible under either ordering—but the paper cannot currently be reproduced from its equations, and the interpretation of Figure 5 as 'appropriate complexity' is ambiguous. We do not see a need to move away from the reader's conditional verdict; a clarification or code release would settle the concern.","tokens_in":15395,"tokens_out":7952,"duration_ms":77658,"concrete_test":"Obtain the released code (or request it from the authors) and print the actual r_i and p_i per expert for the MoCE-IR configuration. Verify whether p_i is increasing or decreasing in i, and compare with the formula r = C/2^i. Then recompute Figure 5 routing patterns: check whether the experts selected for SOTS, Rain100L, and BSD68 have large p or small p, and whether the per-expert FLOPs at 224x224 (with the actual w_i, r_i, and FFT-WSA) are ordered consistently with p_i. If p_i follows the printed formula (expert 1 largest), the paper's description of the mechanism is wrong; if it follows the text (expert 1 lightest), Eq. 3.1 should be corrected to r = C/2^{n-i+1} or equivalent, and the authors should state which schedule was used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 3.1, the expert embedding is defined as r = C/2^i for i in {1,...,n}, giving r1 = C/2 > r2 = C/4 > ... > rn = C/2^n. Yet the same paragraph states 'the most lightweight expert is characterized by the smallest embedding dimension r1 and window size w1'. Since projection layers scale with r, this makes expert 1 the parameter-heaviest, not the lightest. Equation 6 computes the bias b_i = p_i/p_max, and the text says this 'assigns progressively lower weights to experts with lower complexity'. If p_i decreases with i, then b_i decreases with i, so the auxiliary loss (Eq. 7) will favor high-index experts (small r, large window w). That is not a bias toward lower complexity if complexity is identified with parameter count or FLOPs. The paper never defines a single complexity axis: channel width and window size are traded off, so p_i and FLOPs need not be monotonic in the same direction. The central claim is that this bias drives task-specific allocation, and Tab. 5b shows the bias is necessary; an inconsistent definition of the expert ordering therefore makes the mechanism ambiguous. If the intended schedule is r = C/2^{n-i+1}, the equations and text can be reconciled, but then the routing patterns and efficiency story may differ from the printed formula; if the printed formula is what is implemented, the bias is reversed relative to the description.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MoCE-IR, an all-in-one image restoration model in which each decoder block contains a mixture-of-experts layer with experts of different channel widths and window sizes, together with a shared expert. A complexity-aware auxiliary loss biases routing toward lower-complexity experts, with the stated goal of making task-specific allocation emerge automatically and allowing irrelevant experts to be bypassed at inference. The authors validate the method on three-degradation, five-degradation, and composited-degradation benchmarks, report gains over AirNet, PromptIR, InstructIR, OneRestore, and other baselines at lower parameter counts, and provide ablations on expert scaling, routing bias, routing visualizations, and fixed-expert analysis.","tokens_in":15756,"tokens_out":8545,"duration_ms":79774,"significance":"The empirical scope is a strength: evaluations cover AIO-3, AIO-5, and CDD11 composite settings at two model scales, with ablations isolating the complexity bias (Tab. 5b), routing visualizations (Fig. 5), and a fixed-expert test (Tab. 6). The code and models are promised publicly, and the efficiency measurements in Tab. 4 directly address the practical motivation of bypassing experts. If the central mechanism is as described, the paper would make a useful contribution by showing that a simple parameter-count bias in the load-balancing loss induces task-specific routing without explicit degradation labels. However, the current manuscript contains an internal contradiction in the expert complexity schedule that affects the interpretation of the claimed bias, and the quantitative claims would benefit from repeated-run statistics.","major_comments":[{"comment":"The expert-complexity ordering is internally inconsistent. The text says the most lightweight expert has the smallest embedding dimension r1 and window size w1, but Eq. (2) defines r = C/2^i, so for i=1,...,n expert 1 has the largest embedding (r1 = C/2) and expert n the smallest (rn = C/2^n). Because the projection layers in Eq. (2) are linear in r, parameter count p_i in Eq. (6) decreases with i, which reverses the labels used in the routing discussion and in Fig. 5, where the y-axis is labeled 'increasing expert complexity'. Moreover, parameter count and receptive field are traded off (r decreases while w increases), so the bias b_i = p_i/p_max is not an unambiguous complexity proxy. The authors must state the implemented schedule (e.g., r_i = C/2^{n-i+1}), verify it against the released code, and either justify p_i as the complexity measure or report per-expert FLOPs.","section":"Sec. 3.1, Eq. (2); Sec. 3.2, Eq. (6)"},{"comment":"All quantitative comparisons are reported as single PSNR/SSIM values with no repeated runs or confidence intervals. Several of the claimed gains are small (e.g., 0.02-0.04 dB over UniProcessor in Tab. 1 and 0.03 dB over InstructIR in Tab. 2), so without variance estimates it is not possible to judge whether the differences are significant. Please add at least three seeds for the main tables or report the training variance, and clarify how the mean and standard deviation of FLOPS/runtime in Tab. 4 were computed (which tasks, how many runs, and whether router overhead is included).","section":"Sec. 4, Tables 1-4"},{"comment":"The fixed-expert analysis that supports the task-discriminative claim is incomplete. Table 6 reports only rain and noise, omitting haze, and the learned routing for noise (E4, 33.92 dB) is actually slightly worse than the manual choice E4 (34.00 dB), which weakens the claim that learned routing finds the best expert. The rows for the load-balancing baseline are marked 'Not Applicable' without explanation. Please extend the analysis to all degradations and report the full routing matrix per layer/task, or temper the conclusion that the method 'automatically' assigns every task to its most suitable expert.","section":"Sec. 4.2, Table 6 and Fig. 5"}],"minor_comments":[{"comment":"Equation (2) and the implementation-details paragraph both write 'i ∈ {i, ..., n}'; this should be 'i ∈ {1, ..., n}'.","section":"Sec. 3.1"},{"comment":"There is a typo in 'Withing(x), we select...'; it should be 'Within g(x), ...'.","section":"Sec. 3.2"},{"comment":"The load-balancing baseline is cited as [45] in Tables 5b and 6, but reference [45] is an augmented-reality article; the load-balancing loss in the text is attributed to Riquelme et al. [43], so the table citations should be corrected.","section":"Tables 5b and 6"},{"comment":"The exact definitions of 'nested' and 'exponential' expert scaling, and of the '2(2+i)' window progression, are not given; please provide the precise formulas.","section":"Sec. 4.2, Tab. 5a"},{"comment":"The caption should define what is averaged in the routing heatmaps (images, tokens, or layers) and specify the color scale units.","section":"Fig. 5"},{"comment":"The introduction contains a typo: 'Noteable works' should be 'Notable works'.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the empirical breadth is solid, but the contradiction between Eq. (2) and the complexity-ordering text is central to the paper's story and must be resolved before acceptance. Please ask the authors to state the exact implemented r_i schedule and to confirm it against the released code. The single-run results are also a concern given the small margins in Tables 1 and 2. Finally, the acknowledgements thanking reviewers/program committee and the citation of [45] for load balancing should be cleaned up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the complexity-expert idea is real and the paper mostly delivers, but the central scaling formula in Sec. 3.1 is written in a way that contradicts the surrounding text, and that needs to be sorted out before the routing story can be trusted.\n\nWhat is actually new: replacing uniform experts with a set of blocks that trade channel width against window size, and adding a parameter-count bias to the router that pushes inputs toward cheaper experts. That combination is not in Riquelme's uniform MoE or in the degradation-prior MoEs (WM-MoE, AMIR). The paper does a decent job of showing the bias matters: Tab. 5b gives 32.39 with standard load balancing vs. 32.57 with the complexity bias, and Fig. 5 / Tab. 6 show task-dependent expert allocation rather than uniform usage. The evaluation is broad, covering 3, 5, and composited degradations, and the efficiency numbers include runtime variance. The fixed-expert analysis is a nice sanity check.\n\nSoft spots, in order:\n1. The expert scaling is inconsistent as written. Equation 2 and the implementation text say r = C/2^i, which makes expert 1 the widest (r = C/2) and expert n the narrowest. The prose says the most lightweight expert has the smallest r1 and w1, and says dimensionality is progressively reduced in subsequent experts. Those cannot both be true. Since the routing bias in Eq. 6 is built on per-expert parameter counts, the direction of the bias depends on which schedule was actually used. If the printed formula is the real one, the bias is effectively reversed relative to the description. This is a one-line fix (index reversal), but as submitted it is a load-bearing ambiguity, not a cosmetic typo.\n2. No error bars on PSNR/SSIM. Several wins over strong baselines are 0.1–0.6 dB; without repeated runs or significance info, the SOTA claims are weaker than the tables look.\n3. “Any Image Restoration” overstates the scope: three, five, and composited degradations is good coverage but not any.\n4. The delta over the authors' own prior MoE restoration works [62, 63] should be stated explicitly; right now it reads as an incremental step, which is fine, but the paper should own it.\n\nThe stress-test note is correct: the formula/text mismatch is real and should be fixed. The rest of the empirical story looks solid. I'd send this to peer review rather than desk reject; a referee can ask for the corrected scaling description and error bars. The paper is worth engaging, and the routing visualization is a useful contribution.","headline":"Complexity-scaled experts with a parameter-count routing bias is a genuinely new idea, and the experiments largely support it, but the expert-scaling equation contradicts the text and needs fixing before the routing story can be fully trusted.","tokens_in":16295,"tokens_out":7102,"would_cite":true,"duration_ms":67930,"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":"A cheap-first routing bias makes mixture-of-experts restoration networks task-specialized and state-of-the-art.","keywords":["all-in-one image restoration","mixture-of-experts","complexity experts","task-discriminative routing","image dehazing","deraining","denoising","dynamic networks"],"falsifier":"Train the same MoCE layer in the AIO-3 setting twice, once with the parameter-count bias and once with a bias computed from measured per-expert FLOPs or latency, and compare both routing histograms and average PSNR: if the task-to-expert mapping and the reported 32.57 dB average are unchanged, parameter count is an adequate complexity proxy, while if the mapping shifts or quality drops, the paper's causal story fails.","tokens_in":15227,"feed_emoji":"🖼️","tokens_out":8458,"duration_ms":73730,"temperature":0.7,"pith_summary":"The paper introduces MoCE-IR, a single network for all-in-one image restoration, and argues that the key to making mixture-of-experts work there is to give the experts different computational complexity and receptive fields, then bias routing toward the cheaper ones. The authors' central claim is that this cheap-first bias, without any task labels, drives task-specific allocation: localized degradations such as rain land on small experts, while haze and noise, which need broader context, land on larger ones. They report state-of-the-art PSNR/SSIM on three, five, and composited degradation benchmarks at a lower average FLOP count than comparable models. A sympathetic reader should care because the recipe is nearly parameter-free, one auxiliary-loss bias computed from parameter counts, yet the experiments show it changes routing behavior qualitatively and improves restoration quality.","feed_headline":"Cheap-first bias makes image-restoration experts specialize per task","feed_subtitle":"MoCE-IR routes haze, rain, and noise to matched-complexity experts, beating prior all-in-one models with fewer FLOPs.","key_machinery":"The load-bearing mechanism is the mixture-of-complexity-experts (MoCE) layer, made of $n$ nested expert blocks whose embedding dimension shrinks as $r_i = C/2^i$ and whose window size grows, plus a shared transposed self-attention expert. A top-1 router picks one expert for the whole image under noisy softmax routing, and an auxiliary loss combines the coefficient of variation of expert importance and load, with importance multiplied by the complexity bias $b = [p_1/p_{\\max}, \\ldots, p_n/p_{\\max}]$ based on each expert's learnable-parameter count. This bias is the only mechanism steering the router toward simpler experts, and the paper's ablations show it carries the task-discriminative behavior.","core_discovery":"The central discovery is that a mixture-of-experts layer built from non-uniform complexity experts, blocks with progressively smaller channel width and larger window partitions, spontaneously learns task-discriminative routing when the router's importance loss is scaled by a complexity bias $b = [p_1/p_{\\max}, \\ldots, p_n/p_{\\max}]$. With image-level top-1 routing, the model assigns haze and denoising to high-capacity experts, rain to a lightweight local expert, and lets some experts serve several degradations at once. The paper shows this specialization is caused by the bias: replacing it with standard load balancing dissolves the specialization and lowers average PSNR from 32.57 dB to 32.30 dB on the three-degradation setting.","pith_inferences":["A testable extension not explored here: replace the parameter-count bias with a FLOPs-based or latency-based bias and check whether the same task-to-expert mapping survives.","An implication the authors do not draw: because the router learns task identity without labels, the trained routing weights could serve as a lightweight degradation classifier.","A possible next step the paper mentions but does not test: patch-level routing for a single image containing multiple degradations, which image-level routing cannot handle.","If the cheap-first principle is general, inserting MoCE-style experts into other restoration backbones should reproduce the specialization, which would make the recipe portable."],"forward_implications":["Inference can bypass high-complexity experts on easy inputs: the light model averages 36.93 GFLOPS and the heavy model 80.59 GFLOPS, with runtime around 22–23 ms at 224×224 on an RTX 4090.","The same network wins or ties the best previous results across all three benchmark groups: AIO-3 average PSNR 32.57 dB, AIO-5 average 30.58 dB, and CDD11 average 29.05 dB (light model), while using fewer parameters than most of its rivals.","Standard load balancing in place of the complexity bias drops AIO-3 average PSNR from 32.57 dB to 32.30 dB, which the paper presents as evidence that the bias, not the non-uniform experts alone, drives the gain.","Fixed post-training expert assignments show that complexity-biased routing produces larger performance differences between experts for a given task than load-balanced routing, so the router is genuinely task-discriminative."],"supporting_citations":[{"why":"Defines the all-in-one restoration setting and datasets (SOTS, Rain100L, BSD68) and serves as the baseline MoCE-IR must beat.","marker":"[24]"},{"why":"Provides the training configuration and the PromptIR baseline whose prompt-based route is compared against complexity routing.","marker":"[36]"},{"why":"Supplies the importance/load-balancing auxiliary loss that the complexity bias modifies and the standard-MoE baseline in ablations.","marker":"[43]"},{"why":"Provides the Restormer backbone and the transposed self-attention module used as the shared expert.","marker":"[65]"},{"why":"Establishes the five-degradation (AIO-5) evaluation protocol and the IDR baseline.","marker":"[66]"},{"why":"Supplies the FFT-based self-attention approximation used inside each complexity expert block.","marker":"[21]"},{"why":"Contributes the CDD11 composited-degradation benchmark and the OneRestore baseline for the eleven-setting comparison.","marker":"[17]"}],"fun_headline_variants":["Complexity bias makes MoE experts specialize per task","Task-discriminative experts from a cheap-first routing rule","MoCE-IR: complexity-based expert routing beats uniform MoE","Simple complexity bias yields task-specialized MoE experts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The mechanism depends on parameter count per expert being a faithful proxy for computational complexity and for the receptive field a degradation needs, because only that count enters the routing bias.","fun_headline_variants_meta":{"raw":{"variants":["Complexity bias makes MoE experts specialize per task","Task-discriminative experts from a cheap-first routing rule","MoCE-IR: complexity-based expert routing beats uniform MoE","Simple complexity bias yields task-specialized MoE experts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1518,"prompt_tokens":942,"completion_tokens":576,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":509}},"tokens_in":558,"tokens_out":576,"duration_ms":5749,"temperature":1.0,"reasoning_tokens":509,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:10:52.824322+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same MoCE layer in the AIO-3 setting twice, once with the parameter-count bias and once with a bias computed from measured per-expert FLOPs or latency, and compare both routing histograms and average PSNR: if the task-to-expert mapping and the reported 32.57 dB average are unchanged, parameter count is an adequate complexity proxy, while if the mapping shifts or quality drops, the paper's causal story fails.","supporting_citations":[{"cited_title":"All-In-One Image Restoration for Unknown Corruption","cited_arxiv_id":null,"evidence_quote":"Defines the all-in-one restoration setting and datasets (SOTS, Rain100L, BSD68) and serves as the baseline MoCE-IR must beat."},{"cited_title":"Promptir: Prompting for all-in-one image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the training configuration and the PromptIR baseline whose prompt-based route is compared against complexity routing."},{"cited_title":"Scaling vision with sparse mixture of experts","cited_arxiv_id":null,"evidence_quote":"Supplies the importance/load-balancing auxiliary loss that the complexity bias modifies and the standard-MoE baseline in ablations."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the Restormer backbone and the transposed self-attention module used as the shared expert."},{"cited_title":"Ingredient-oriented multi- degradation learning for image restoration","cited_arxiv_id":null,"evidence_quote":"Establishes the five-degradation (AIO-5) evaluation protocol and the IDR baseline."},{"cited_title":"Efficient frequency domain-based trans- formers for high-quality image deblurring","cited_arxiv_id":null,"evidence_quote":"Supplies the FFT-based self-attention approximation used inside each complexity expert block."},{"cited_title":"Onerestore: A universal restoration framework for com- posite degradation","cited_arxiv_id":null,"evidence_quote":"Contributes the CDD11 composited-degradation benchmark and the OneRestore baseline for the eleven-setting comparison."}],"review_version":1}