{"id":"d85b2835-972a-438f-9290-4eb0ab71f45c","arxiv_id":"2506.21722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A diffusion-based training framework with matched time steps improves generalization and unified multi-task performance of image restoration networks.","lead":"This paper shows how to train ordinary image restoration networks with diffusion-style generative objectives so they generalize better to unseen degradation. The method adds generative pre-training, matched time steps, and incremental multi-task training, improving reported performance with no extra inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training-set-derived matching time step tmat is the load-bearing premise; its stochastic computation and target-set validation leave transferability unestablished.","rationale":"The paper's central claim has two empirical pillars: single-task generalization gains (Setting A) and multi-task unified gains (Setting B). Both pillars are mediated by the matching time step tmat. In Setting A, tmat determines the fine-tuning time step and the layer-wise decay wdecay(t); in Setting B, tmat fixes the incremental training order and the MoE prompt. If tmat is not a stable, source-computable property of the degradation, then the method as described is not a general recipe but a benchmark-tuned procedure, which would directly weaken the title-level claim of a general IR training paradigm. The reader's weakest_assumption identified exactly this premise, and I agree that it is the most load-bearing assumption. The concern is not that tmat is definitely wrong; the paper may well have computed it faithfully on training data. Rather, the manuscript does not provide the evidence needed to rule out the alternative that the reported gains come from selecting t on the target evaluation sets. The proposed test resolves this by comparing source-only tmat to source-validation and target-oracle choices. A secondary issue is the internal inconsistency between main-text Table 3 and supplementary Table 10: the Setting B ablation values 17.76 and 27.01 are attached to GEF in one table and to TIT in the other. This does not by itself falsify the central claim, but it compounds the difficulty of attributing the multi-task gain to a specific mechanism, and it should be corrected. Because the concern is testable and the core evidence is substantial, the appropriate verdict remains CONDITIONAL; if the proposed test shows tmat transfers, the concern is resolved, and if it fails, the framework's generality claim would need to be substantially revised.","tokens_in":28123,"tokens_out":11064,"duration_ms":121700,"concrete_test":"Fix one degradation (e.g., dehazing on RESIDE-to-REVIDE). Compute tmat on the RESIDE training split via Algorithm 2 over 10 random seeds, and again on a randomly held-out subset of RESIDE. Then fine-tune the generative-pretrained network under three conditions: (a) t = tmat computed from training data only, (b) t chosen to maximize PSNR on a held-out source-domain validation split, and (c) t chosen to maximize PSNR on REVIDE (the oracle used implicitly in Fig. 3d). If condition (a) underperforms (c) by a margin comparable to the reported generalization gain, or if tmat varies by more than a few steps across seeds or data subsets, the transferability premise fails. Separately, rerun the Setting B ablations with shuffled incremental-task order and with the TIT/GEF labels from Table 3 and Table 10 disambiguated to resolve the swapped 17.76/27.01 entries.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that tmat, computed in Algorithm 2 as the step minimizing ||(xt-xt-1)-(y-x)|| on training pairs, is a stable property of each degradation type and transfers to (i) the fine-tuning time step in GEF and (ii) the ordering of incremental training and the time-based MoE prompts in MTUL. This premise is insecure for three reasons. First, Algorithm 2 line 9 is under-specified: the minimization is over neither samples nor random realizations, and xt-xt-1 depends on the sampled noise z, so the returned tmat is stochastic and no variance or seed-level stability is reported. Second, the only evidence that tmat matches the best fine-tuning step comes from Fig. 3(b,d), where time steps are swept and evaluated on the target generalization sets; this does not show that tmat computed from source training data alone would select the same step without access to the target. Third, Table 4 lists tmat values for ten degradations (4, 8, 12, 15, 19, 22, 27, 38, 47, 50) with no error bars, no cross-dataset replication, and no comparison to other possible orderings. If tmat is a training-set artifact or is effectively chosen on the benchmark, the proposed matching mechanism, the GEF time-step conditioning, and the TIT ordering lose their stated basis, and the reported generalization and multi-task gains may not transfer to new degradation types or datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training framework that transplants the diffusion generative training paradigm into general reconstruction-based image restoration (IR) networks. The method has three stages: Timestep Matched Generative Pre-training (TMGP) on clean images, Generalization Enhanced Fine-tuning (GEF) that mixes generative and reconstructive objectives with regularization losses, and Multi-task Unified Learning (MTUL) that uses a degradation-specific matching time step tmat to order incremental training and to condition MoE adapters. The authors report substantial single-task generalization improvements (e.g., dehazing generalization PSNR 20.74 vs. best prior 17.26 in Table 1) and state-of-the-art average multi-task performance on 10 degradations (Table 2, average PSNR 28.13 vs. RAM's 27.51), with ablations supporting each component and extensions to NAFNet and X-Restormer.","tokens_in":28392,"tokens_out":4528,"duration_ms":54386,"significance":"The paper's core idea—using the diffusion time-step structure as an organizing principle for IR training—is interesting and potentially useful for the community, especially because the framework is architecture-agnostic and the authors provide pseudo-code, extensive ablations, visualizations, and complexity comparisons. The strengths of the manuscript include the systematic exploratory analysis in Section 2, the breadth of experiments on 14 datasets, and the attempt to unify single-task generalization and multi-task unified IR under one training paradigm. However, the central premise that tmat is a stable, transferable property of a degradation type is not convincingly established, and the multi-task comparison protocol may not be equitable. If these issues are resolved, the framework could be a valuable training recipe for general IR networks.","major_comments":[{"comment":"The matching time step tmat is the load-bearing quantity of the framework, but its computation is under-specified and its transferability is not demonstrated. Algorithm 2 line 9 writes 'tmat ← t If min ∥(xt − xt−1) − (y − x)∥²' without specifying whether the minimization is over time steps, over samples in the training set, or over random realizations of the reverse-process noise z; since xt−1 depends on the sampled noise z in line 7, the returned tmat is stochastic, yet Table 4 reports single values with no variance or seed-level stability. More importantly, Fig. 3(b,d) sweeps the fine-tuning time step and evaluates on the target generalization sets (REVIDE, Rain100H, RealSnow, LOL-v2); this only shows that some time step chosen with access to the target data improves generalization, not that tmat computed from source training data alone would select that same step. Because tmat also determines the TIT ordering and the MoE prompts in Section 3.3, the entire framework's advantage rests on a premise that is currently unsupported. Please report tmat distributions over samples/seeds, verify that source-computed tmat matches the target-sweep optimum without target-set access, and compare TIT against alternative orderings (e.g., reverse, random, or complexity-based).","section":"Section 3.1, Algorithm 2, Table 4, and Fig. 3(b,d)"},{"comment":"The claim of superior multi-task unified IR performance requires that the Setting B baselines are trained on the same 10-task mixture, the same data splits, and a comparable training budget. The manuscript does not state whether the numbers for AirNet, Restormer, NAFNet, PromptIR, IR-SDE, DA-CLIP, ResShift, X-Restormer, and RAM are taken from their original papers or retrained under the authors' protocol. The data mixture described in Appendix G.1 is non-standard (e.g., a subset of Snow100K-L, CelebA-HQ for inpainting with RePaint masks, and a specific low-light set), so published numbers trained on different mixtures are not directly comparable. Please either retrain all baselines under the identical mixture and budget or clearly report the source of each baseline number and provide a matched-training comparison; otherwise the headline average PSNR gain of 0.62 dB over RAM could be an artifact of training-set composition rather than the proposed method.","section":"Section 4, Table 2, and Appendix G.1"},{"comment":"The parameter importance regularization Lreg is not implementable as written. Equations (3)-(5) define Ωθk in terms of the post-fine-tuning parameters θ1, but the regularization loss is supposed to be used during fine-tuning before those parameters exist; the notation is also inconsistent (θ1_r vs. θ0_k vs. θ1_k). Algorithm 6 appears to compute δθ0_k = θ1_k − θ0_k only at the 'last training epoch' and then performs a single update θ1_k ← θ0_k + ε∇Ls, which does not correspond to normal gradient-descent training and would not produce the reported results. The ablation tables (Tables 6-10) show that Lreg contributes meaningfully to performance, so this component cannot remain a black box. Please provide a correct, implementable algorithm (e.g., an EWC-style importance computed at the end of pre-training and fixed during fine-tuning) and clarify how the Gauss-Newton approximation in Eq. (25) is used in practice.","section":"Appendix F, Algorithm 6, and Eqs. (3)-(5)"}],"minor_comments":[{"comment":"Fig. 4(b) and Fig. 5 appear to show the same 'matching t guided incremental training' plots with nearly identical captions; please remove the duplicate or differentiate the two figures.","section":"Section 2.1, Fig. 4 and Fig. 5"},{"comment":"The main text refers to 'Eq. 17', 'Eq. 18', and 'Eq. 19' for the gradient orthogonality terms, but these equations appear earlier in the main text as Eqs. (6)-(8); the appendix renumbers them as (17)-(19). Please use consistent equation numbering and correct all in-text references.","section":"Throughout, equation numbering"},{"comment":"The condition 'If min ∥(xt − xt−1) − (y − x)∥²' is formally incomplete because the domain of minimization is not stated; specify whether the minimum is taken over t, over training samples, or over the sampled noise z.","section":"Algorithm 2, line 9"},{"comment":"The hyperparameter notation is inconsistent: the main text says 'λ = 0.2' while Appendix G.2 says 'α = 0.2'; please unify the notation.","section":"Section G.2"},{"comment":"Table 4 reports tmat values without error bars or statistics, and Fig. 3(c) presents claims about optimal network layers without numerical values or error bars; adding quantitative support would strengthen the conclusions.","section":"Table 4 and Fig. 3(c)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid experimental core, but the central tmat transferability issue and the unclear baseline protocol in Setting B are exactly the kinds of concerns that will be raised by reviewers. The parameter-importance pseudo-code is, as written, not reproducible and needs a careful rewrite. I would encourage the editor to request a revision focused on these three points rather than rejecting, because the framework is novel and the ablations are otherwise fairly consistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before reading. First, this is a genuine framework paper: the authors combine diffusion generative pre-training with reconstruction-based IR training, and the empirical gains are substantial (dehazing generalization PSNR 20.74 vs 17.26 for X-Restormer; multi-task average 28.13 vs 27.51 for RAM). Second, the load-bearing component — the degradation-specific matching time step tmat — is not convincingly shown to transfer from source training data to target generalization. That is the paper's soft underbelly, not the experiments.\n\nWhat is actually new: the combination of generative pre-training on clean images, degradation-matched time-step selection, fine-tuning with parameter-importance regularization and gradient orthogonality, time-sequential incremental training, and MoE adapters with time-based prompts. The exploratory analysis in Sec. 2 (time-step dependencies, layer-wise sensitivity, mixing ratios) is genuinely useful, and the ablations are extensive and mostly consistent. The method can be bolted onto existing backbones, and the inference-time overhead is essentially zero. That is a practically useful result.\n\nThe soft spots. First, the tmat premise: Algorithm 2 is under-specified — the minimization is over neither samples nor random realizations, and xt-1 depends on sampled noise z, so tmat is stochastic. No variance or seed-level stability is reported. More importantly, the only evidence that tmat selects the best fine-tuning step comes from sweeps evaluated on the target generalization sets (Fig. 3b,d). That does not establish that source-data-only tmat would pick the same step without access to the target. Table 4 lists values without error bars or cross-dataset replication. If tmat is a training-set artifact or effectively chosen on the benchmark, the stated basis for GEF time-step conditioning and TIT ordering collapses. Second, baseline fairness in Setting B is unclear: the paper does not state whether the baselines were trained on the same 10-task mixture. That should be clarified. Third, the theoretical derivation of the importance regularizer (Eq. 3-5 and the appendix version) is mathematically sloppy — notation issues and a hand-wavy Taylor expansion. The regularizer empirically works, so this is fixable, but the derivation does not support the paper's framing. Fourth, no code or data is released, which is a reproducibility concern.\n\nOverall, the central claim is plausible and testable, and the paper deserves a serious referee. The reviewer should push hard on tmat transferability and baseline fairness.","headline":"A real framework with big empirical gains, but the matching time step tmat is a load-bearing premise whose transferability is not established.","tokens_in":28962,"tokens_out":2382,"would_cite":true,"duration_ms":24472,"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":"Transplanting the diffusion generative training objective into ordinary reconstruction-based restoration networks, guided by a per-degradation matching time step, improves generalization to unseen degradations and unifies ten restoration…","keywords":["image restoration","diffusion models","generative pre-training","generalization","multi-task unified IR","matching time step","incremental training","parameter regularization"],"falsifier":"Compute the matching time step $t_{\\mathrm{mat}}$ on one training split of a degradation (say RESIDE), then fine-tune the pre-trained network at $t_{\\mathrm{mat}}$ and at neighboring time steps and evaluate on an out-of-distribution target (REVIDE); if $t_{\\mathrm{mat}}$ does not coincide with the time step achieving the best target-domain PSNR, the framework's central premise fails. A cheaper check: re-estimate $t_{\\mathrm{mat}}$ on a second training split or with a different training seed; if the estimated $t_{\\mathrm{mat}}$ shifts by more than a few steps, it cannot be a stable degradation property.","tokens_in":27897,"feed_emoji":"🖼️","tokens_out":7919,"duration_ms":79853,"temperature":0.7,"pith_summary":"The paper sets out to prove that the diffusion training paradigm—the objective and schedule used to train generative diffusion models—can be grafted onto ordinary, reconstruction-based image restoration networks without changing their architecture or adding iterative sampling at inference. It claims that generative pre-training on clean images, followed by fine-tuning at a degradation-specific matching time step with regularization that preserves generative knowledge, sharply improves generalization to unseen degradation distributions in single-task restoration. For unified multi-task restoration, it claims that ordering incremental data introduction by the same matching time steps and adding time-prompted MoE adapters yields a single model that beats prior all-in-one methods across ten degradation types. If these claims hold, existing lightweight restoration backbones could inherit the robustness of diffusion models while keeping fast single-pass inference.","feed_headline":"Diffusion pretraining lifts image restoration generalization","feed_subtitle":"A per-degradation matching time step guides fine-tuning and unifies ten restoration tasks in one model.","key_machinery":"The central organizing device is the degradation-specific matching time step $t_{\\mathrm{mat}}$: for each degradation type, the pre-trained generative network runs the reverse diffusion process on a degraded image, and $t_{\\mathrm{mat}}$ is the time step at which the successive-difference $(\\mathbf{x}_t - \\mathbf{x}_{t-1})$ best matches the degradation residual $(\\mathbf{y} - \\mathbf{x})$. This single scalar summarizes degradation complexity, selects the fine-tuning time step, orders the incremental multi-task training schedule, and supplies the time-based prompts to the MoE adapters. Around it the framework adds three regularizers: second-order parameter-importance regularization $L_{\\mathrm{reg}}$ that slows forgetting of generative knowledge during fine-tuning, a gradient-orthogonality loss $L_{\\mathrm{orthog}}$ that aligns generative and reconstruction gradient directions, and a layer-wise weight decay $w_{\\mathrm{decay}}(t)=e^{-at}$ that confines generative fine-tuning to shallow, domain-invariant layers.","core_discovery":"Diffusion generative pre-training on clean images gives an ordinary reconstruction-based restoration network $r_\\theta(\\mathbf{x}, t)$ access to the diffusion latent space, and each degradation type is assigned a matching time step $t_{\\mathrm{mat}}$ via reverse sampling such that the reversed image aligns with the clean image. Fine-tuning the pre-trained network at that $t_{\\mathrm{mat}}$, mixing about 10% generative data into the reconstruction fine-tuning, constraining updates with parameter-importance regularization and gradient orthogonality, and applying generative updates only to shallow layers produces restoration networks whose out-of-distribution generalization exceeds existing backbones in four single-task settings and whose unified model outperforms prior methods across ten degradations.","pith_inferences":["If $t_{\\mathrm{mat}}$ is indeed a stable property of each degradation type, it offers a cheap proxy for degradation complexity: one could predict relative transfer difficulty between degradations and schedule data collection or fine-tuning for new degradations without retraining, a program the paper does not itself pursue.","The shallow-layer/generative versus deep-layer/reconstruction split suggests a general recipe for injecting generative priors into discriminator-style networks, which may extend beyond restoration to super-resolution, JPEG artifact removal, or compressive sensing.","The paper reports $t_{\\mathrm{mat}}$ increasing with global, complex degradations (low-light $t_{\\mathrm{mat}}=38$, haze $t_{\\mathrm{mat}}=47$, blur $t_{\\mathrm{mat}}=50$), hinting at a monotone degradation-complexity axis; a testable extension is to verify this ordering on graded synthetic degradations.","The paper's stated limitation is extra training time; a natural follow-up is distilling the generative pre-training into fewer epochs or merging it into the fine-tuning stage, and the framework's own ablations suggest $L_{\\mathrm{reg}}$ and $L_{\\mathrm{orthog}}$ carry much of the benefit."],"forward_implications":["In single-task settings, the framework raises out-of-distribution PSNR by large margins, e.g., dehazing trained on RESIDE reaches 20.74 dB on REVIDE versus 17.26 dB for the best compared prior.","In unified multi-task settings, the framework reaches 28.13 dB average PSNR across ten degradations, outperforming the 27.51 dB of the best prior, with the best LPIPS among compared methods.","The framework is backbone-agnostic: applying it to NAFNet and X-Restormer lifts their unified-restoration averages to 28.47 and 28.32 dB respectively.","A mixing ratio of roughly 10% generative data during fine-tuning maximizes the benefit; larger generative ratios begin to degrade restoration fidelity.","Ordering incremental multi-task training by matching time steps gives better unified performance than directly mixing all task data."],"supporting_citations":[{"why":"Supplies the DDPM forward and reverse processes and the generative training objective $L_\\gamma$ used in Timestep Matched Generative Pre-training.","marker":"[10]"},{"why":"Provides the score-based SDE view of diffusion that motivates treating the latent space as a source of degradation-to-clean mappings and the matching-time-step analysis.","marker":"[51]"},{"why":"Basis for the second-order parameter-importance regularization $L_{\\mathrm{reg}}$ that preserves generative knowledge during restoration fine-tuning.","marker":"[14]"},{"why":"Gives the accelerated $T=50$ schedule and the reverse-sampling update used in Algorithm 2 when computing $t_{\\mathrm{mat}}$.","marker":"[46]"},{"why":"Main prior baseline in both Setting A and Setting B; the paper follows its training settings and benchmarks against it.","marker":"[26]"},{"why":"Strongest prior on the ten-task unified benchmark with average PSNR 27.51, which the proposed framework surpasses at 28.13.","marker":"[36]"},{"why":"Serves as a general restoration backbone that the framework extends, reaching 28.32 average PSNR when integrated.","marker":"[5]"},{"why":"Provides the NAFNet backbone used to demonstrate that the framework lifts an existing lightweight restoration network to 28.47 average PSNR.","marker":"[3]"},{"why":"Recent diffusion-as-loss approach that the paper distinguishes from its combined generative-and-reconstruction objective.","marker":"[17]"}],"fun_headline_variants":["Diffusion training paradigm boosts general image restoration","Diffusion pretraining with matched time steps unifies ten restorations","Regularized diffusion pretraining for unified image restoration","Diffusion latent access improves multi-task restoration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's gains ride on the assumption that the degradation-specific matching time step $t_{\\mathrm{mat}}$, computed on training data, is a stable property of the degradation type that transfers to choosing fine-tuning time steps and ordering incremental training; if $t_{\\mathrm{mat}}$ is instead a training-set artifact or needs tuning on the target generalization data, the advantage collapses.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion training paradigm boosts general image restoration","Diffusion pretraining with matched time steps unifies ten restorations","Regularized diffusion pretraining for unified image restoration","Diffusion latent access improves multi-task restoration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000397,"raw_usage":{"total_tokens":2045,"prompt_tokens":880,"completion_tokens":1165,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1103}},"tokens_in":496,"tokens_out":1165,"duration_ms":10212,"temperature":1.0,"reasoning_tokens":1103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:20:05.173938+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the matching time step $t_{\\mathrm{mat}}$ on one training split of a degradation (say RESIDE), then fine-tune the pre-trained network at $t_{\\mathrm{mat}}$ and at neighboring time steps and evaluate on an out-of-distribution target (REVIDE); if $t_{\\mathrm{mat}}$ does not coincide with the time step achieving the best target-domain PSNR, the framework's central premise fails. A cheaper check: re-estimate $t_{\\mathrm{mat}}$ on a second training split or with a different training seed; if the estimated $t_{\\mathrm{mat}}$ shifts by more than a few steps, it cannot be a stable degradation property.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM forward and reverse processes and the generative training objective $L_\\gamma$ used in Timestep Matched Generative Pre-training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the score-based SDE view of diffusion that motivates treating the latent space as a source of degradation-to-clean mappings and the matching-time-step analysis."},{"cited_title":"Kirkpatrick, R","cited_arxiv_id":null,"evidence_quote":"Basis for the second-order parameter-importance regularization $L_{\\mathrm{reg}}$ that preserves generative knowledge during restoration fine-tuning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the accelerated $T=50$ schedule and the reverse-sampling update used in Algorithm 2 when computing $t_{\\mathrm{mat}}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Main prior baseline in both Setting A and Setting B; the paper follows its training settings and benchmarks against it."},{"cited_title":"Restore Anything with Masks: Leveraging Mask Image Modeling for Blind All-in-One Image Restoration","cited_arxiv_id":"2409.19403","evidence_quote":"Strongest prior on the ten-task unified benchmark with average PSNR 27.51, which the proposed framework surpasses at 28.13."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a general restoration backbone that the framework extends, reaching 28.32 average PSNR when integrated."}],"review_version":1}