{"id":"c8ac1f9b-76b8-4df9-80fb-bf4196e1c8ae","arxiv_id":"2608.09482","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MGN-AIR, a pixel-level multimodal guidance network for all-in-one image restoration, reports state-of-the-art average PSNR on several benchmarks, including a 1.51 dB gain on the CDD11 composite degradation benchmark.","lead":"The authors propose a new image restoration network that uses per-pixel maps to guide where and how much to restore, instead of applying one global fix to the whole image. They report large gains on composite degradations such as rain plus haze, but the paper lacks code and shows weaker results on pure denoising.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on pixel-level guidance whose supervision target (Eq. 4) conflates image content with degradation severity; if the visual prompt does not track true per-pixel degradation, the reported gains are not attributable to the proposed mechanism.","rationale":"The reader's weakest assumption identifies the same load-bearing point: Eq. (4) assumes the pixel-wise absolute difference between clean and degraded images is a faithful per-pixel degradation map. I agree that this is the most consequential unexamined assumption. The novelty of MGN-AIR is explicitly 'pixel-level restoration driven by local degradation intensity'; if the supervision target for that intensity is miscalibrated, then the proposed mechanism is not actually doing what the paper claims, even if the reported PSNR numbers are reproducible. The strength of the paper is its architecture and the substantial reported gains on composite degradation, and those gains could in principle come from the auxiliary loss, added capacity, or frequency-domain attention rather than from meaningful pixel-level guidance. That is why I do not move the verdict to reject: the empirical claim is not disproven, and the paper has plausible ablations. But the conditional accept should require either code release or a direct evaluation of the predicted visual prompt against the Eq. (4) target, plus an ablation removing L_Aux, before the mechanistic claim is accepted. I also note the paper's 'consistently outperforms' wording is contradicted by its own Table 2 and Table 3 on denoising, but that overclaim is secondary to the VPGM concern because the paper explicitly acknowledges the denoising regression in Section 4.2.","tokens_in":16059,"tokens_out":5787,"duration_ms":69811,"concrete_test":"Retrain MGN-AIR on CDD11 with the auxiliary visual-prompt loss removed (L_Aux = 0) while keeping all other components identical, and compare the resulting average PSNR with the full model. Additionally, on the held-out CDD11 test set, compute the Spearman correlation between the VPGM output V and the ground-truth map X-hat from Eq. (4); if the correlation is low (e.g., below 0.3) or if removing L_Aux does not reduce the CDD11 gain substantially, the claimed pixel-level guidance is not the source of the improvement. For a stronger diagnostic, train a variant that is given the oracle Eq. (4) map as the visual prompt at both training and inference; if oracle guidance does not outperform the learned VPGM, the degradation-map formulation itself is not providing the advertised benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — 1.51 dB average PSNR gain on CDD11 and SOTA across benchmarks — is only as strong as the VPGM's ability to estimate per-pixel degradation severity. Section 3.2 defines the supervision target in Eq. (4) as the channel-averaged, min-max normalized absolute difference |I_HQ - I_LQ|. This quantity is not a faithful degradation map. For low-light and haze, the absolute difference is dominated by the clean image's local intensity: bright structures produce large differences even when the degradation is spatially uniform, so the 'degradation map' encodes content, not corruption level. For Gaussian noise, the target is a realization of the noise field, which cannot be inferred from the degraded image alone; the VPGM is trained to predict an irreducible random component. The paper's ablation (Table 4) shows a large drop when the visual prompt is removed, and this is interpreted as evidence that fine-grained pixel-level guidance is the mechanism. But the drop could equally come from the auxiliary loss L_Aux, the added FFT-attention module, or the extra parameters in the MGB. Since the paper releases no code and reports no accuracy metric for the predicted visual prompt itself, the headline improvement on CDD11 cannot be independently attributed to pixel-level guidance. The method may still be a strong empirical system, but the specific mechanistic claim that makes it novel is unsupported without evidence that the learned visual prompt actually tracks per-pixel degradation intensity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MGN-AIR, an all-in-one image restoration network that performs restoration at the pixel level by combining a visual prompt, estimated by a Visual Prompt Generation Module (VPGM), with a CLIP-based textual prompt in a Multimodal Guidance Module (MGM). The resulting restoration matrix is used in a Pixel-Level Restoration Module (PLRM) to blend local/dilated convolutions with spatial attention per pixel. The method is evaluated on the CDD11 composite-degradation benchmark and on standard five- and three-degradation all-in-one settings, reporting average PSNR/SSIM improvements over prior methods, including a 1.51 dB average PSNR gain over MoCE-IR on CDD11, together with ablations and a complexity comparison.","tokens_in":16349,"tokens_out":7857,"duration_ms":78817,"significance":"If the pixel-level guidance mechanism performs as claimed, the work would be a meaningful advance over global prompt-based all-in-one restoration methods, and the reported CDD11 improvement is substantial. Strengths of the manuscript include the breadth of the evaluation, the explicit auxiliary supervision of the visual prompt, and the complexity-control experiments in Table 5. However, the central mechanistic claim is not yet sufficiently supported: the supervision target in Eq. (4) conflates image content with degradation severity, the ablation in Table 4 is confounded, and the own tables contradict the phrase 'consistently outperforms' on the denoising task. These issues are load-bearing for the main contribution, so a major revision is required.","major_comments":[{"comment":"The supervision target for the visual prompt is the channel-averaged, min-max normalized absolute difference between the clean image and the degraded image. This quantity does not faithfully represent per-pixel degradation severity: for additive Gaussian noise it is a realization of the noise field and is not identifiable from the degraded image alone, while for low-light and haze it is dominated by the clean image's local intensity and edges. The paper provides no evaluation of the predicted visual prompt itself (e.g., correlation with actual degradation severity), so the claimed pixel-level guidance mechanism is not established. Please add such an evaluation or adopt a more physically motivated supervision target.","section":"§3.2, Eq. (4)"},{"comment":"The ablation 'w/o visual prompt' removes the VPGM, the auxiliary loss L_Aux, the FFT-attention branch, and the associated parameters together, so the observed performance drop cannot be attributed specifically to pixel-level guidance. A control experiment with matched capacity and auxiliary supervision but with a content-agnostic or global prompt is needed, together with a quantitative accuracy metric for the predicted visual prompt, before the 0.73 dB gap over 'w/o MGB' can be assigned to the proposed mechanism.","section":"§4.2, Table 4"},{"comment":"The abstract's claim that the method 'consistently and significantly outperforms' existing approaches is contradicted by the paper's own denoising numbers: on BSD68 at sigma=25, MGN-AIR reports 31.12 dB versus DFPIR's 31.29 dB in Table 2, and the averaged denoising PSNR in Table 3 is 31.00 dB versus DFPIR's 31.29 dB and PromptIR's 31.12 dB. The claim should be qualified to average performance or accompanied by an explanation of this task-specific trade-off.","section":"§4.2, Tables 2 and 3"},{"comment":"The comparison protocol is ambiguous: DA-CLIP is listed as a compared method in the text but is absent from Table 2, and the paper does not state whether the baseline numbers were obtained by retraining under the same data, patch sizes, and optimizer settings or were taken from the original papers. Without this information, the state-of-the-art claim on the five-degradation benchmark is not fully verifiable.","section":"§4.1, §4.2, Table 2"}],"minor_comments":[{"comment":"The text refers to 'MoCE-IR [56]' in the CDD11 comparison and in the Fig. 3 caption, but reference [56] is DFPIR; the correct citation for MoCE-IR is [70].","section":"§4.2 and Fig. 3 caption"},{"comment":"The notation '·' in Eq. (3) is ambiguous; if it denotes elementwise multiplication in the Fourier domain, the operation implements circular convolution rather than the dot-product attention described in the text. Please clarify the exact operation and how the Softmax is applied.","section":"§3.2, Eq. (3)"},{"comment":"The dimensions of T_p and the role of the scalar beta in 'Concat[Avg(T); beta]' are unclear; please specify how the 1D convolution transforms the concatenated vector into a spatial feature.","section":"§3.3, Eq. (5)"},{"comment":"The text says 'Ours-s' keeps a similar parameter count to PromptIR, but it has 26.61M versus 32.97M parameters while using more FLOPs (133.42 versus 121.08 G) and higher latency (35.84 versus 31.09 ms); the claim should be phrased more carefully.","section":"§4.2, Table 5"},{"comment":"The paper does not release code or trained models; for a method whose main contribution is a new mechanism, providing code or at least predicted visual prompt maps would substantially strengthen reproducibility and help verify the mechanism.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of ACM MM and presents an interesting idea with strong average benchmark numbers. The main risk is that the attribution of the improvements to pixel-level guidance is under-supported and the comparison protocol is not fully specified. I would ask the authors to add a direct evaluation of the visual prompt, a deconfounded ablation, and a clarification of the baseline protocol before considering acceptance. Also check the [56]/[70] citation mix-up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper could be an important data point for all-in-one restoration if the numbers hold, but the mechanism it credits for the gains is not supported by the evidence as written. Worth having a referee look at it, but the authors need to do real work on the attribution and the overclaims.\n\nWhat's new: per-pixel visual prompt learned from the degraded image, combined with a CLIP text prompt to produce a restoration matrix that switches between local and global operations per pixel. That's a sensible step beyond global prompt-based methods like PromptIR and DA-CLIP. The CDD11 composite degradation results are remarkably strong, 1.51 dB over MoCE-IR on average, and the method also does well on the five-degradation and three-degradation benchmarks.\n\nThe architecture is clean and the ablations are the right shape: removing the visual prompt hurts, removing the text prompt hurts, and the lightweight variant still outperforms baselines.\n\nNow the soft spots, in proportion.\n\nThe central attribution problem is real. The visual prompt is supervised by Eq. (4), the channel-averaged, min-max normalized absolute difference between clean and degraded images. That quantity is not a degradation map. For Gaussian noise it is a random field; for low-light it is dominated by clean-image intensity; for haze it is content-dependent. The VPGM cannot reliably predict it from the degraded image alone, so the claimed per-pixel degradation signal may be largely content-correlated or near-constant. The big ablation drop when the visual prompt is removed is suggestive, but it does not isolate the prompt: it also removes the auxiliary loss, the FFT-attention module, and some parameters. The paper reports no metric for how well the visual prompt matches the actual degradation, so the mechanism is not demonstrated.\n\nThe \"consistently and significantly outperforms\" claim is contradicted by the paper's own Table 2: denoising 31.12 dB vs DFPIR's 31.29 and MoCE-IR's 31.34, and low-light 23.67 vs DFPIR's 23.82. The average is better, but consistency is false.\n\nThere are also presentation issues: DA-CLIP is mentioned in the text but missing from Table 2, there are citation numbering errors, and no code or training details are released. Also, the paper never says how the text prompt is chosen at inference; if it needs the ground-truth degradation type, that should be explicit.\n\nTo be fair, the circularity concern is low: the final restoration is not derived from the map by a fitted derivation, the map is just an auxiliary supervision. That part I don't see as a problem.\n\nWho this is for: anyone working on prompt-based all-in-one restoration, especially for mixed degradations. They should read it, but with care.\n\nMy recommendation: do not desk-reject. Send it to peer review, but the reviewers should require the authors to show that the visual prompt actually tracks degradation severity, to clarify the text-prompt setting, and to correct the overclaims. If the results are real, this is a useful paper after revision.","headline":"Large reported gains on composite degradation benchmarks, but the pixel-level guidance mechanism is not actually demonstrated; the paper overclaims consistency.","tokens_in":16895,"tokens_out":5215,"would_cite":false,"duration_ms":53745,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that all-in-one image restoration improves when guidance is per-pixel rather than global, and reports a 1.51 dB PSNR gain over the previous best on composite degradations.","keywords":["all-in-one image restoration","pixel-level guidance","visual prompt","textual prompt","degradation map","composite degradation","image denoising","low-light enhancement"],"falsifier":"Take clean images, add no degradation, and compute the Eq. (4) map: if the map is not nearly uniform and close to zero but instead highlights edges or bright textures, then the supervision is measuring content rather than degradation. Alternatively, add uniform noise to a flat region and leave a high-texture region clean; if the model assigns higher severity to the clean textured region than to the noisy flat region, the pixel-level guidance is miscalibrated.","tokens_in":15830,"feed_emoji":"🖼️","tokens_out":7027,"duration_ms":63491,"temperature":0.7,"pith_summary":"The paper argues that all-in-one image restoration models should not treat an image uniformly, because real degradations are spatially uneven: rain hits streaks, haze varies by depth, and noise may be localized. It proposes MGN-AIR, which predicts a per-pixel degradation map, combines it with a text description of the degradation type, and uses the resulting restoration matrix to choose a different restoration operation at each pixel. On the CDD11 benchmark, where several degradations co-occur, the method reports a 1.51 dB PSNR improvement over the nearest recent approach, and it also leads on a five-task all-in-one benchmark. The largest gains appear exactly where the pixel-level idea should help: images mixing multiple degradation types.","feed_headline":"Pixel-level guidance beats uniform restoration by 1.51 dB","feed_subtitle":"MGN-AIR estimates a per-pixel degradation map to guide restoration, outperforming prior all-in-one models.","key_machinery":"The load-bearing object is the pixel-level restoration matrix $P\\in[0,1]^{H\\times W\\times C}$, produced by the Multimodal Guidance Module from the visual prompt $V$, a per-pixel degradation severity map, and a global textual prompt. $P$ gates two restoration paths in the Pixel-Level Restoration Module: heavily degraded pixels are restored with standard and dilated convolutions, while lightly degraded pixels go through spatial attention over an $8\\times 8$ window. The visual prompt itself is trained against a prescribed degradation map, Eq. (4), given by the min-max normalized per-pixel absolute difference between the clean and degraded images averaged across channels.","core_discovery":"MGN-AIR's central claim is that degradation-adaptive all-in-one restoration should be controlled at pixel resolution, not image level. The network's Visual Prompt Generation Module predicts a per-pixel degradation map, supervised at the first encoder level by the min-max normalized per-pixel absolute difference between clean and degraded images averaged over channels. A Multimodal Guidance Module fuses this map with a frozen text encoder's global degradation-type embedding to produce a restoration matrix $P\\in[0,1]^{H\\times W\\times C}$. A Pixel-Level Restoration Module uses $P$ to blend a local convolution path, for heavily degraded pixels where only neighbors can be trusted, with a spatial-attention path, for lightly degraded pixels where recurring structural patterns elsewhere help. The paper reports consistent gains over prior all-in-one methods across three-, five-, and eleven-task benchmarks, with the largest margin on the composited CDD11 set.","pith_inferences":["The same pixel-level design could transfer to real-world restoration by replacing the supervised degradation map with an unsupervised severity estimate, since paired clean-degraded images are rarely available in practice.","The Eq. (4) supervision ties the visual prompt to any large pixel difference, so gains on synthetic benchmarks may shrink on real photographs where edges and illumination changes also produce large differences.","The predicted restoration matrix could serve as an editable control: users could raise or lower restoration strength in specific regions without retraining.","A natural extension is to supervise the visual prompt at deeper levels as well, which the current design only does implicitly."],"forward_implications":["A single MGN-AIR model handles denoising, deraining, deblurring, dehazing, desnowing, and low-light enhancement without task-specific adaptation.","On the CDD11 composite benchmark, the method reports an average PSNR of 30.56 dB, 1.51 dB above the nearest recent method.","Ablations show that removing the visual prompt lowers performance more than removing the textual prompt, so local intensity cues are the larger source of gain.","The restoration matrix makes per-pixel decisions explicit, so the model's choice of local versus global restoration can be inspected at each pixel."],"supporting_citations":[{"why":"Supplies the global visual prompt baseline and the U-Net encoder-decoder architecture that MGN-AIR extends.","marker":"[47]"},{"why":"Introduces the CDD11 composite degradation benchmark and the OneRestore baseline that MGN-AIR outperforms by 2.09 dB on average PSNR.","marker":"[17]"},{"why":"Provides the recent multi-expert baseline MoCE-IR that MGN-AIR surpasses by 1.51 dB on CDD11.","marker":"[70]"},{"why":"Provides the degradation-aware perturbation baseline DFPIR that MGN-AIR improves on in the five-degradation setting.","marker":"[56]"},{"why":"Supplies the classic all-in-one AirNet baseline and a comparison point for architecture and complexity.","marker":"[25]"},{"why":"Represents the human-instruction textual prompt baseline InstructIR that MGN-AIR compares against.","marker":"[8]"},{"why":"Represents the multimodal prompt baseline DA-CLIP that motivates combining text and image cues.","marker":"[37]"},{"why":"Supplies the frozen text encoder that turns the degradation-type phrase into the global embedding used in the multimodal guidance module.","marker":"[49]"}],"fun_headline_variants":["Per-pixel guidance refines all-in-one restoration","MGN-AIR: pixel-level multimodal prompts beat uniform","All-in-one restoration gets per-pixel precision","Pixel-aware restoration: MGN-AIR outperforms prior models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the min-max normalized per-pixel absolute difference between clean and degraded images is a faithful measure of degradation severity at each pixel, so strong edges, textures, or illumination differences in image content could mislead the visual prompt.","fun_headline_variants_meta":{"raw":{"variants":["Per-pixel guidance refines all-in-one restoration","MGN-AIR: pixel-level multimodal prompts beat uniform","All-in-one restoration gets per-pixel precision","Pixel-aware restoration: MGN-AIR outperforms prior models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000475,"raw_usage":{"total_tokens":2354,"prompt_tokens":937,"completion_tokens":1417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1352}},"tokens_in":553,"tokens_out":1417,"duration_ms":14370,"temperature":1.0,"reasoning_tokens":1352,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:44:45.503981+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take clean images, add no degradation, and compute the Eq. (4) map: if the map is not nearly uniform and close to zero but instead highlights edges or bright textures, then the supervision is measuring content rather than degradation. Alternatively, add uniform noise to a flat region and leave a high-texture region clean; if the model assigns higher severity to the clean textured region than to the noisy flat region, the pixel-level guidance is miscalibrated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the global visual prompt baseline and the U-Net encoder-decoder architecture that MGN-AIR extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the CDD11 composite degradation benchmark and the OneRestore baseline that MGN-AIR outperforms by 2.09 dB on average PSNR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the recent multi-expert baseline MoCE-IR that MGN-AIR surpasses by 1.51 dB on CDD11."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the degradation-aware perturbation baseline DFPIR that MGN-AIR improves on in the five-degradation setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classic all-in-one AirNet baseline and a comparison point for architecture and complexity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the human-instruction textual prompt baseline InstructIR that MGN-AIR compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the frozen text encoder that turns the degradation-type phrase into the global embedding used in the multimodal guidance module."}],"review_version":1}