{"id":"ce3d8261-e1cb-4a37-a632-571fad369eb5","arxiv_id":"2608.06939","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A prompt-based restoration network that couples textual degradation descriptions with visual features reports the best average PSNR and SSIM on four all-in-one weather removal benchmarks.","lead":"This paper introduces DCMPC-Net, a restoration network that turns text descriptions of weather from a frozen vision-language model into degradation-aware prompts and injects them into a transformer backbone to clean rain, fog, snow, and raindrops. On four standard benchmarks it reports the best average PSNR and SSIM, about 0.27 dB above the previous best.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed cross-modal semantic contribution is not isolated: no ablation varies text content independently of the visual branch, and the full model adds an FFT loss the baselines do not use; a constant-text control is needed before prompt semantics can be credited with the SOTA margin.","rationale":"The empirical claim 'best average PSNR/SSIM' is checkable from the released code and from Table I, and I do not accuse the numbers of being false. The conceptual claim that degradation-aware semantic prompts cause the gains is the load-bearing part, and it is currently underdetermined. The ablation w/o CMPG (Table IV) shows the whole system minus the prompt generator still exceeds Histoformer on Snow100K-S, so other modules and the added FFT loss already contribute. The CMPG-specific gain is concentrated on Outdoor-Rain and Snow-L; on RainDrop the gain is 0.02 dB, within run-to-run noise. More importantly, Table V's text-features variant performs worse than Histoformer on Outdoor-Rain and RainDrop, which is hard to reconcile with a story in which text semantics are the key signal. The full CMPG could succeed because the visual branch supplies degradation features and the text branch acts as a fixed query/key regularizer. Table VI's learnable-query ablation underperforms, but a constant-text control would distinguish 'frozen text embedding helps' from 'the content of the description matters.' Because Eqs. (4)-(5) are dimensionally inconsistent, the paper itself cannot be used to verify the mechanism; the code is the arbiter. The existing CONDITIONAL verdict is appropriate; I would keep it and make acceptance contingent on this constant-text control, and ideally an FFT-loss ablation.","tokens_in":22187,"tokens_out":13873,"duration_ms":126664,"concrete_test":"Using the released DCMPC-Net code, keep all weights, the CMPG architecture, the cached-prompt protocol, and the loss in Eq. (13) fixed, but replace every LLaMA-derived description T_text used in Eq. (1) with one identical string (e.g., 'an image') for all training and test images, so that T_emb is constant across the dataset. Recompute PSNR/SSIM on Snow100K-S, Snow100K-L, Outdoor-Rain, and RainDrop. If the average is within 0.05 dB of the reported 34.03 dB / 0.9486, the semantic content of the descriptions is not load-bearing; if it drops materially, the degradation-aware text premise is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C explicitly adds an FFT loss 'in addition to those used in the baseline setting,' and no ablation removes it, so the 0.27 dB average margin over CyclicPrompt (Table I) is not attributable solely to the proposed modules. The prompt-specific evidence is also incomplete. In Table V, replacing CMPG with text features alone gives 31.95 dB on Outdoor-Rain and 32.96 dB on RainDrop, below the Histoformer backbone (32.08 and 33.06), so the text branch is not beneficial in isolation. In Table VI, learnable queries are worse than the frozen-text branch, but this only shows that a frozen text initialization helps; it does not show that the weather-specific content of the descriptions is the cause. If a single constant text description performs equally in the cached-prompt protocol, then P_da is effectively a visual prompt with a fixed text query, and the 'degradation-aware semantic' explanation is not load-bearing. This concern is compounded by the dimension errors in Eqs. (4)-(5): T_emb is L×C and I_emb is H×W×C, yet Q, K, V are asserted to be H×W×C, and P_da = V·softmax(K·Q/β) is not a valid cross-attention contraction. The formal description is therefore not self-contained, and the released code is the only precise specification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DCMPC-Net, an all-in-one adverse-weather restoration network built on the Histoformer backbone, augmented by a Cross-Modal Prompt Generator (CMPG) that uses a frozen vision-language model and text encoder to produce degradation-aware prompts, plus a Prompt-Guided Attention Alignment Module (PGAAM) and a Dual Feature Compensation Module (DFCM). The method is evaluated on Snow100K-S/L, OutdoorRain, and RainDrop, reporting the best average PSNR/SSIM (34.03 dB / 0.9486) among the compared methods, with ablations over the proposed components, semantic sources, and module placements.","tokens_in":22493,"tokens_out":9211,"duration_ms":74651,"significance":"If the central attribution is correct, the paper makes a useful empirical contribution by showing that frozen VLM-derived semantic prompts can be injected as feature-level conditioning in a unified restoration backbone, and the code release and explicit failure-case discussion are positive features. The average gain of 0.27 dB over the second-best method is modest but non-negligible. However, the evidence for the degradation-aware semantic mechanism is not yet isolated: an extra FFT loss is used without ablation, a constant-text control is missing, the unimodal text branch is sometimes worse than the backbone, and the formal equations for the core cross-attention are dimensionally invalid. These issues must be addressed before the specific semantic claim can be accepted.","major_comments":[{"comment":"Equations (4)-(5) are not algebraically consistent with the stated tensor shapes. With T_emb ∈ R^{L×C} and I_emb ∈ R^{H×W×C}, the projections cannot all produce Q, K, V ∈ R^{H×W×C}; moreover, P_da = V·softmax(K·Q/β) is not a valid cross-attention contraction because K·Q does not have compatible dimensions (no transposition is indicated, and the text sequence length L is not spatial H×W). Since CMPG is the central novelty, the paper is not self-contained without corrected equations or an explicit pointer to the exact implementation.","section":"Section III-B, Eqs. (4)-(5)"},{"comment":"Section IV-C states that the FFT loss is incorporated 'in addition to those used in the baseline setting,' but no experiment removes L_fre. Consequently, the 0.27 dB average margin over CyclicPrompt in Table I, and the improvements of the full model over Histoformer in Table V, cannot be attributed solely to CMPG/PGAAM/DFCM. Please add an ablation with and without L_fre under otherwise identical settings, and ensure that compared baselines are trained with the same loss.","section":"Section IV-C and IV-E, Eq. (13)"},{"comment":"The current ablations do not isolate the weather-specific semantic content of the LLaMA descriptions. In Table V, replacing CMPG with text features alone yields 31.95 dB on OutdoorRain and 32.96 dB on RainDrop, below the Histoformer backbone (32.08 and 33.06), so unimodal text guidance is not beneficial without the rest of the pipeline. In Table VI, learnable queries are worse than the frozen-text branch, but this only shows that a frozen text initialization helps. A control experiment in which CMPG receives a fixed, semantically vacuous text description (e.g., 'a photo') with all other components and losses unchanged is needed to determine whether the degradation-specific wording matters. Without this control, the 'degradation-aware semantic' explanation is not load-bearing.","section":"Section IV-E.2, Tables V-VI"},{"comment":"The paper claims to 'outperform state-of-the-art methods' in the abstract and Section I, but Table I shows DCMPC-Net does not achieve the best per-dataset PSNR on OutdoorRain (32.49 vs 32.81 for CyclicPrompt) or on Snow100K-L (32.35 vs 32.37 for T3-DiffWeather). Table II similarly shows LDR has higher PSNR on Snow100K-L (32.47 vs 32.35) and RainDrop (33.83 vs 33.08). The supported claim is best average PSNR/SSIM, not consistent superiority; please rephrase the claims to match the evidence.","section":"Section IV-D.1, Tables I-II"},{"comment":"Table VII reports Histoformer Snow100K-S SSIM as 0.9689, while Tables I and V report 0.9658 for the same configuration. This changes the apparent margin of the full model (0.9692 vs 0.9658 = +0.0034 in Table I, but +0.0003 against 0.9689 in Table VII) and complicates the ablation interpretation. The table also groups 'BLIP & CLIP' into one row despite the text saying results are derived from BLIP, CLIP, or no-prompt configurations; please report these variants separately and correct the inconsistent number.","section":"Section IV-E.2, Table VII"},{"comment":"The paper describes LLaMA [35] as a 'pre-trained vision-language model' and writes T_text = PVL(Q_t, I_LQ), feeding the input image to LLaMA. LLaMA is a language model and cannot process image inputs unless a multimodal adapter is used (e.g., LLaVA). Please specify the exact VLM and image-input interface, and correct the related statement that it was pre-trained on 15 trillion multimodal tokens.","section":"Section III-B.1, Eq. (1)"}],"minor_comments":[{"comment":"The header 'Histormer' is a typo for 'Histoformer'.","section":"Table I"},{"comment":"Please state which losses constitute the baseline setting so that the FFT-loss addition is precisely defined.","section":"Section IV-C"},{"comment":"Please define F(·) as the Fourier transform operator and specify the normalization used in the frequency-domain loss.","section":"Section III-E, Eq. (16)"},{"comment":"The variables y_t and y_{t+1} appear in the DFCM equations and in Fig. 4 but are not formally introduced in the text; please define them.","section":"Section III-D, Eq. (12) and Fig. 4"},{"comment":"Many configurations share identical Snow100K-S PSNR/SSIM values (e.g., 38.00/0.9683), which may reflect rounding at 0.01 dB; consider reporting more significant digits to make the ablations informative.","section":"Table VI"},{"comment":"The row label 'BLIP & CLIP' is ambiguous; please report BLIP and CLIP separately, along with the prompt template used for each.","section":"Table VII"},{"comment":"For comparisons with LDR computed from the official repository, please confirm that the same evaluation protocol (Y-channel PSNR/SSIM, border cropping, and full-resolution restoration) was used as for the other methods.","section":"Section IV-D.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript has a credible empirical core and provides code, but the main semantic-attribution claim currently rests on ablations that do not control for the added loss or for text content. The dimensional errors in Eqs. (4)-(5) and the LLaMA/multimodal confusion are fixable but make the method section not self-contained. I recommend major revision rather than rejection; the authors should be asked to add the constant-text control and FFT-loss ablation, correct the equations and tables, and tone down the per-dataset SOTA claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid all-in-one weather removal paper with a real architectural idea: rather than using language semantics only as an external description or a router (LDR), it fuses frozen LLaMA/E5 text embeddings with visual features into a degradation-aware cross-modal prompt, then conditions the decoder through an attention-alignment module and a dual compensation branch. That combination is new relative to the cited LDR, DA-CLIP, and PromptIR lines. The empirical side is thorough—four benchmarks, module ablations, a user study, a downstream detection example, and honest reporting of the 9B-parameter full-pipeline cost under the cached-prompt protocol. The code link is a plus.\n\nI agree with your conditional verdict, with one correction: Histoformer is not the authors' own backbone. It is Sun et al. (ECCV 2024), so using it as baseline and in ablations is standard practice, not a circularity problem.\n\nThe soft spots the stress test identifies are real. First, Section IV-C says the FFT loss is added \"in addition to those used in the baseline setting,\" and there is no ablation removing it. That alone prevents the 0.27 dB average margin over CyclicPrompt from being attributed to the proposed modules. A one-line ablation dropping L_fre, or adding the same loss to CyclicPrompt, would fix this.\n\nSecond, the semantic-contribution claim is under-supported. Table V shows replacing CMPG with text features alone drops OutdoorRain and RainDrop below the Histoformer backbone, so the text branch is not beneficial in isolation. The comparisons to learnable queries and PromptIR blocks show the frozen-text protocol helps, but they do not vary text content. A constant-text control—say, \"describe the image\" or random tokens—is needed before \"degradation-aware semantics\" can be credited. Right now, the evidence is consistent with a fixed textual initialization that largely acts as a regularized visual prompt.\n\nThe dimension errors in Eqs. (4)–(5) are genuine: a text embedding of length L cannot be projected to an H×W×C query as written, and P_da = V·softmax(K·Q/β) is not a valid cross-attention contraction. The released code may be the real spec, but the formal description should stand alone.\n\nTwo smaller issues: the abstract overstates the wins—DCMPC-Net loses to CyclicPrompt on OutdoorRain PSNR and to T3-DiffWeather on Snow100K-L PSNR, so it is \"best average,\" not uniformly better. And Table VII lists Histoformer Snow100K-S SSIM as 0.9689, which contradicts Tables I and V (0.9658).\n\nWho it's for: researchers in prompt-based restoration and all-in-one weather removal will want to know this result and the cached-prompt protocol. It deserves a serious referee. I would send it out, but with major revision clearly in mind—the FFT ablation and constant-text control are necessary before the headline claim is solid.","headline":"Solid all-in-one weather removal with a plausible cross-modal prompt design, but the semantic story is under-supported and the added FFT loss muddies the headline margin.","tokens_in":23018,"tokens_out":5784,"would_cite":false,"duration_ms":46926,"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":"This paper proposes DCMPC-Net, an all-in-one adverse-weather removal network whose degradation-aware cross-modal prompts—built by cross-attending frozen language-model text embeddings with image features—outperform prior task-specific and…","keywords":["adverse weather removal","all-in-one image restoration","cross-modal prompt learning","vision-language model","degradation-aware prompts","attention alignment","snow removal","raindrop removal"],"falsifier":"Run the full model with the text descriptions replaced by deliberately generic or wrong captions (for example 'clear sunny scene' for every rainy input) while keeping all modules identical; if the average PSNR does not fall toward the no-prompt baseline, then the cross-modal semantic content is not what drives the gain. Alternatively, compute prompts live at test time instead of loading cached embeddings and check whether accuracy drops, which would reveal whether the cached-prompt protocol is masking the online cost of semantic extraction.","tokens_in":22015,"feed_emoji":"🌧️","tokens_out":5266,"duration_ms":44319,"temperature":0.7,"pith_summary":"The paper proposes DCMPC-Net, a single network for removing rain, snow, fog, and raindrops that conditions a shared restoration backbone with degradation-aware prompts built from both text and image cues. Its central claim is that coupling language-derived semantic descriptions of weather with visual features, then injecting the resulting prompts into the decoder through attention alignment, systematically improves restoration accuracy over both the backbone alone and prior prompt-based all-in-one methods. The reported result is an average 34.03 dB PSNR and 0.9486 SSIM across Snow100K-S, Snow100K-L, Outdoor-Rain, and RainDrop, 0.27 dB above the second-best compared method. If true, the paper shows that semantic knowledge from a frozen vision-language model can be made spatially useful for low-level restoration without changing the computation path or routing experts.","feed_headline":"Cross-modal prompt network tops four weather benchmarks at 34.03 dB","feed_subtitle":"Text-and-image prompts let one model beat prior all-in-one methods on four weather benchmarks by 0.27 dB.","key_machinery":"The central object is the Cross-Modal Prompt Generator (CMPG), which produces a degradation-aware cross-modal prompt $P_{da}$ by cross-attention between a frozen text encoder's embedding $T_{emb}$ (query) and visual features $I_{emb}$ (key and value): $P_{da} = V \\cdot \\mathrm{Softmax}(K \\cdot Q / \\beta)$. This latent prompt is injected into the decoder through the Prompt-Guided Attention Alignment Module (PGAAM), whose dual cross-attention establishes spatial correspondence between semantic cues and degraded regions; the Dual Feature Compensation Module (DFCM) additionally uses the channel-wise max-minus-min of the input as a compensation map to reinforce structures. The prompt is the load-bearing mechanism that carries semantic degradation knowledge into feature refinement.","core_discovery":"DCMPC-Net discovers that degradation-aware cross-modal prompts—latent representations formed by cross-attending text embeddings from a frozen language model with degraded image features—can serve as effective feature-level conditioning signals in a unified adverse-weather removal network. These prompts are aligned to degraded regions by a Prompt-Guided Attention Alignment Module, while a Dual Feature Compensation Module separates degradation artifacts from scene structure to preserve fine detail. The paper argues this design outperforms image-level visual prompts, text-only cues, and routing-based language priors because it retains spatial correspondence and provides continuous modulation rather than discrete expert selection. The strongest quantitative claim is the best average PSNR and SSIM across the four benchmarks, with the largest gains on Outdoor-Rain where degradation is spatially uneven.","pith_inferences":["Our inference: if the frozen model's descriptions become generic or misaligned with local degradation, the cross-modal prompt should degrade toward a static token; this is testable by ablating the prompt query across levels of vagueness and predicting a monotonic drop in the Outdoor-Rain gain.","Our inference: the same conditioning mechanism could transfer to other spatially heterogeneous degradations such as reflections, shadows, or local blur, where text describes the artifact type but attention must localize it; the paper does not test these settings.","Our inference: the cached-prompt protocol means real-time cost is decoupled from the frozen model, so a natural next step—not explored here—is learning a lightweight prompt estimator that replaces the offline extraction pass."],"forward_implications":["One trained model can handle snow, rain&fog, and raindrops together, outperforming task-specific and previous all-in-one methods on the four evaluated benchmarks.","Language-derived semantics can be injected as continuous feature-conditioning signals without expert routing or explicit degradation maps, keeping the restoration computation path fixed.","Each component contributes additively: removing the cross-modal prompt generator costs about 0.2 dB on Outdoor-Rain, showing that prompt semantics matter at the feature level rather than as a mere auxiliary description.","Because prompts are precomputed and cached, the expensive frozen semantic-prior stage can run offline, leaving the restoration-stage runtime at 0.08 seconds per 256x256 image.","The ablation against LDR indicates that converting vision-language priors into a continuous prompt is competitive with converting them into routing scores, with DCMPC-Net reporting higher average SSIM across the three shared test sets."],"supporting_citations":[{"why":"Supplies the histogram-prior transformer backbone that DCMPC-Net conditions with cross-modal prompts and the primary baseline for all ablations.","marker":"[31]"},{"why":"Provides the frozen vision-language model whose textual descriptions of weather type, intensity, and obscured areas carry the semantic degradation cues.","marker":"[35]"},{"why":"Frozen text encoder that maps the language-model descriptions into the 1024-dimensional embeddings used in the cross-modal prompt generator.","marker":"[51]"},{"why":"Defines the All-Weather training set that mixes Snow100K, Outdoor-Rain, and RainDrop pairs for the unified restoration model.","marker":"[16]"},{"why":"The language-driven routing-based restoration method compared directly against DCMPC-Net on three shared test sets.","marker":"[27]"},{"why":"Provides the PromptIR-style learnable prompt block used as a lightweight-prompting alternative in the CMPG ablation.","marker":"[21]"},{"why":"Supplies the Snow100K training and test data, including Snow100K-S and Snow100K-L evaluation sets.","marker":"[56]"},{"why":"Supplies the Outdoor-Rain heavy rain and fog training and test data.","marker":"[57]"},{"why":"Supplies the RainDrop real-world raindrop image pairs used for training and testing.","marker":"[58]"}],"fun_headline_variants":["Cross-modal prompts clear rain, snow, and haze in one net","Text-image prompts let one model beat four weather benchmarks","Degradation-aware prompts improve unified adverse-weather removal","Language cues plus visual prompts boost restoration accuracy across weather","Aligning text prompts with degraded regions yields better weather removal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole gain rests on the frozen vision-language model, prompted with a fixed question, producing text descriptions that genuinely encode the local weather degradation, and on those semantics surviving when the prompts are cached offline and simply loaded during training and testing.","fun_headline_variants_meta":{"raw":{"variants":["Cross-modal prompts clear rain, snow, and haze in one net","Text-image prompts let one model beat four weather benchmarks","Degradation-aware prompts improve unified adverse-weather removal","Language cues plus visual prompts boost restoration accuracy across weather","Aligning text prompts with degraded regions yields better weather removal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2239,"prompt_tokens":958,"completion_tokens":1281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1201}},"tokens_in":574,"tokens_out":1281,"duration_ms":11646,"temperature":1.0,"reasoning_tokens":1201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:12:52.167750+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full model with the text descriptions replaced by deliberately generic or wrong captions (for example 'clear sunny scene' for every rainy input) while keeping all modules identical; if the average PSNR does not fall toward the no-prompt baseline, then the cross-modal semantic content is not what drives the gain. Alternatively, compute prompts live at test time instead of loading cached embeddings and check whether accuracy drops, which would reveal whether the cached-prompt protocol is masking the online cost of semantic extraction.","supporting_citations":[{"cited_title":"All in one bad weather removal using architectural search,","cited_arxiv_id":null,"evidence_quote":"Defines the All-Weather training set that mixes Snow100K, Outdoor-Rain, and RainDrop pairs for the unified restoration model."},{"cited_title":"Language-driven all-in-one adverse weather removal,","cited_arxiv_id":null,"evidence_quote":"The language-driven routing-based restoration method compared directly against DCMPC-Net on three shared test sets."},{"cited_title":"Promptir: Prompting for all-in-one image restoration,","cited_arxiv_id":null,"evidence_quote":"Provides the PromptIR-style learnable prompt block used as a lightweight-prompting alternative in the CMPG ablation."}],"review_version":2}