{"id":"48d77491-edd0-453e-970f-e40f618b85d6","arxiv_id":"2507.22498","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SSGformer, an all-in-one transformer combining Sobel and SVD spectral prompts with mask-based group-wise attention, reports state-of-the-art averages on All-weather and WeatherStream.","lead":"A new neural network, SSGformer, restores photos degraded by rain, snow, and raindrops in one pass by extracting edge and low-frequency clues via Sobel and SVD, then grouping similar image regions for attention. It reports the best average PSNR and SSIM on the All-weather benchmark and top average PSNR on WeatherStream.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The grouping operation that carries the paper's stated novelty is specified only as an even value-based partition; as written this is a hard, non-differentiable assignment, so the paper does not establish that the grouping mask is learned or that the reported gains are attributable to spatial…","rationale":"The reader's weakest assumption is that the grouping-mask reliably captures degradation patterns; my stress-test sharpens this into a trainability and attribution problem. The paper describes an even, hard partition of single-channel mask values, but gives no mechanism for gradients to flow through the partition. The only described gradient path into the mask is the element-wise multiplication in Eq. 3, which trains the mask as a gate. Consequently, the ablations in Tab. 6 (with/without FGA) do not isolate learned spatial grouping; a fixed random mask with the same gating operation might reproduce much of the gain. This concern does not make the empirical benchmark claim false, but it does mean the paper's central architectural explanation is not currently supported by the reported evidence. Since the reader's CONDITIONAL verdict already requires code release, grouping-mechanism specification, and error-bar reporting, and since the proposed checks would resolve exactly these points, I recommend leaving the verdict unchanged rather than moving it.","tokens_in":17090,"tokens_out":5076,"duration_ms":66196,"concrete_test":"Release the training code and check the grouping operator. If it uses argsort/topk/gather without a differentiable surrogate, demonstrate how gradients reach the mask generator through the assignment, or replace it with a differentiable/soft grouping and retrain. In the same setting, run two ablations: (1) freeze the mask generator at random initialization while still applying grouping, and (2) apply a randomly permuted mask; if the PSNR change versus the reported FGA ablation is smaller than the 0.2 dB gap to Histoformer, the learned grouping is not the source of the claimed gains. Finally, sweep gp ∈ {1, 2, 8, 16}; a flat or non-monotonic curve would indicate the specific grouping count is incidental.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SSGformer achieves state-of-the-art all-in-one adverse weather removal, and the paper's explanation for that result is grouping spatially similar pixels using a learned grouping-mask (Sec. 3.3, Sec. 3.4, Eq. 3, Eq. 5). To make that explanation valid, the mask must yield stable, semantically meaningful, and trainable group assignments. The paper specifies only that mask values are 'divided evenly into gp groups' with equal element counts, i.e., a sorting/topk partition of the single-channel mask. Such a hard partition is piecewise constant in the mask values: gradients with respect to the mask vanish almost everywhere through the grouping operation itself. The mask does receive some learning signal from the element-wise multiplication in Eq. 3, (LayerNorm(F'_p) ⊙ Mp), but that path trains the mask as a gating signal, not as a clustering signal. The group-wise attention and the group selector in Eq. 7 involve hard assignments/argmax, and no differentiable-sorting, straight-through estimator, or soft assignment is described. It is therefore possible—and in the absence of code not ruled out—that after training the mask is an arbitrary quantization of a gating feature, and the gains attributed to FGA in Tab. 6 come from the mask-as-gate or from added parameters rather than from grouping similar degradations. This is load-bearing because 'spatial grouping' is the stated novelty; if the grouping is not actually learned, the explanation for the benchmark gains collapses even if the numerical SOTA claim happens to hold for an unidentified reason. The absence of error bars on gains of 0.13–0.35 dB reinforces the need to verify that the mechanism, not chance or extra capacity, drives the reported improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SSGformer, an all-in-one adverse weather removal network. It extracts high-frequency edge features via a Sobel operator and low-frequency features via singular value decomposition, fuses these with multi-head linear attention to produce a degradation-aware feature F_S, and generates a single-channel grouping mask from F_S using a convolution. The mask is then used to spatially group features and to feed a Spatial Grouping Transformer Block (SGTB) with in-group and cross-group attention, where the cross-group partner is selected by a cosine-similarity group selector. The model is evaluated on the All-weather and WeatherStream datasets, reporting the best average PSNR/SSIM on All-weather and WeatherStream, with the strongest margins on rain (+0.35 dB) and raindrop (+0.18 dB). The authors include ablations of the SDP components, attention configuration, group attention, and the combination of FGA and SDP, and they provide a public code link.","tokens_in":17458,"tokens_out":4131,"duration_ms":50472,"significance":"If the spatial grouping mechanism works as described, the paper makes a useful architectural contribution by connecting spectral prompts to spatially grouped attention in multi-weather restoration. The ablation study is reasonably structured and the code is promised, which aids reproducibility. However, the central novelty—the learned grouping mask and the group-wise attention driven by it—is not precisely specified, and the report does not demonstrate that the performance gains are attributable to the grouping itself rather than to the mask-as-gate or to added parameters. The reported margins over the second-best methods are small and are not accompanied by error bars or multiple-seed statistics, so the state-of-the-art claim is not yet rigorously supported.","major_comments":[{"comment":"The grouping operation is underspecified. The text states that features are 'divided evenly into gp groups' and that 'each grouped feature has the same spatial dimensions, i.e., the same number of elements,' but it does not define how the mask values are partitioned (e.g., sorting and equal-size binning), how ties are broken, or how cases where H*W is not divisible by gp are handled. Because the grouping is the core of the claimed contribution, a precise algorithmic definition is required.","section":"Sec. 3.4, Eq. (5) and Fig. 3(b)"},{"comment":"The grouping assignment appears to be a hard, non-differentiable operation. If the mask values are sorted and split into equal-size bins, the assignment is piecewise constant and the gradient with respect to the mask values through the grouping step is zero almost everywhere. The only gradient path to the mask is the element-wise multiplication in Eq. (3), which trains the mask as a gating signal rather than as a clustering signal. The paper does not describe a straight-through estimator, a soft assignment, or a differentiable sorting mechanism, so it is not established that the mask is actually learned to group spatially similar degradation patterns. Please provide a differentiable formulation or an explicit training mechanism, and add an ablation that isolates the grouping assignment from the mask gating (e.g., compare against a fixed random grouping or a fixed spatial partition).","section":"Sec. 3.3 and Eq. (3)"},{"comment":"The reported improvements over the second-best method are small: +0.35 dB on rain, +0.18 dB on raindrop, and +0.13–0.16 dB on WeatherStream sub-tasks. No error bars, standard deviations, or multiple-seed results are reported, so the claim of state-of-the-art performance is not statistically substantiated. Please provide the mean and standard deviation over at least three independent training runs, or a statistical significance test.","section":"Tables 1 and 2"},{"comment":"Several load-bearing hyperparameters are not specified: the number of groups gp per stage (only the example gp=4 is given), the SVD truncation rank used to compute the low-frequency component in the SDP, and the correlation loss weight β in Eq. (10). Without these values, the method is not reproducible, and the sensitivity of the results to gp and the SVD rank is unknown. Please report all hyperparameters and, ideally, a sensitivity analysis for the group count.","section":"Sec. 3 and implementation details"}],"minor_comments":[{"comment":"The phrase 'To address these issue' should be 'To address these issues.'","section":"Abstract"},{"comment":"The header row of Table 2 contains a typo ('PNSR' instead of 'PSNR') and the table formatting appears corrupted; please correct it.","section":"Table 2"},{"comment":"'Specially' should be 'Specifically.'","section":"Sec. 3.2"},{"comment":"The text refers to the supplementary for the objective function details and WeatherStream visualizations, but the supplementary material is not included in this submission. Please ensure it is available to reviewers.","section":"Supplementary"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the overall architecture is coherent. The main risk is that the central 'spatial grouping' mechanism is not precisely defined and its learning signal is questionable; these issues are fixable with a clearer formulation and additional experiments. The small PSNR margins relative to Histoformer also warrant error-bar reporting. I recommend major revision rather than rejection because the authors can address these concerns within the manuscript's current scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the SSGformer paper. It's a competent, incremental contribution to all-in-one adverse weather removal. The genuinely new bit is the combination of Sobel and SVD spectral prompts with a mask-based group-wise attention that includes a learned cross-group selector. That selector, which picks the most similar other group via cosine similarity of representative features, is a reasonable idea and not something I've seen done exactly this way.\n\nThe paper does several things well. The architecture is coherent and the ablations in Tables 3-6 support the contribution of each component. The authors test on both All-Weather and WeatherStream, report consistent gains, and include a limitations section that is honest about the internal-cue strategy. Code is promised. For a subfield that mostly adds capacity or external knowledge, this is a clean, self-contained design.\n\nThe soft spots are mostly about missing specification and evaluation rigor. The grouping rule is described as \"dividing evenly into gp groups\" but the actual operation - sort and split, or threshold? - is never defined. Since it's likely a hard partition, it's unclear how gradients train the mask for grouping rather than as a pure gate. The paper uses the mask in a multiplicative skip-path (Eq. 3), so the mask does receive learning signal as a gate, but the grouping itself is a non-differentiable step. The same issue applies to the argmax in the group selector. This is a genuine specification gap, but I don't think it collapses the paper the way the stress-test suggests. Many grouping methods (Histoformer included) rely on non-differentiable sorting and still work; the mask is input-conditioned, so the partition is a learned function of the image even if the assignment step has no direct gradient. Still, the authors should state this explicitly - either they use a straight-through estimator, differentiable sorting, or they accept that the grouping is a deterministic post-processing of a learned mask.\n\nThe second issue is that the SOTA gains over Histoformer are 0.13-0.35 dB with no error bars or multiple runs. That's a fragile claim at this margin. They should report variance or at least run multiple seeds. The SVD truncation rank, the correlation loss weight, and the group count are also left unspecified.\n\nIn summary, this paper is for image-restoration specialists who want a new grouping-based transformer baseline. It deserves a serious referee - the architecture is novel and the results are plausible - but revision should require a precise description of the grouping and gradient flow, plus error bars.\n\nMy recommendation: send it to review, but with the expectation of significant revision on the method details.","headline":"A solid but incremental all-in-one weather removal architecture with a genuinely new spectral-plus-grouping design, marred by an under-specified grouping rule and small SOTA margins that need error bars.","tokens_in":17988,"tokens_out":3828,"would_cite":false,"duration_ms":45439,"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":"SSGformer removes rain, snow, and raindrops with one network by grouping image regions derived from Sobel edge and SVD low-frequency cues, reporting the best average PSNR on the All-Weather and WeatherStream benchmarks.","keywords":["adverse weather removal","all-in-one image restoration","spectral decomposition","Sobel edge detection","singular value decomposition","group-wise attention","spatial grouping","transformer"],"falsifier":"Replace the learned grouping mask with a fixed random partition of pixels into $g_p$ equal groups and retrain; if All-Weather PSNR stays near 32.63 dB, the mask is not the source of the gain. A cheaper check is to compute the same SSGformer's mask on an image and its horizontally flipped or brightness-shifted version and measure overlap: low overlap would indicate the groups are not stable degradation regions.","tokens_in":16819,"feed_emoji":"🌧️","tokens_out":7788,"duration_ms":84322,"temperature":0.7,"pith_summary":"This paper proposes SSGformer, a single transformer network that restores images degraded by rain, snow, and raindrops. The central claim is that decomposing the degraded image into Sobel-detected edge features and SVD low-frequency components, fusing them, and using the fused features to partition the image into spatially similar groups makes a unified model handle diverse weather distortions better than global frequency filtering. On the All-Weather benchmark the model reports 32.63 dB average PSNR, the best overall result, and on the real-world WeatherStream benchmark it reports 24.77 dB average PSNR. The paper argues that the gain comes from the combination of the spectral prompt and the group-wise attention it enables, not from extra parameters.","feed_headline":"Spectral grouping tops all-in-one weather removal at 32.63 dB PSNR","feed_subtitle":"Sobel edges and SVD low-frequencies guide group-wise attention so one model clears rain, snow, and raindrops.","key_machinery":"The load-bearing mechanism is the grouping-mask $M_p$, a single-channel map produced by a 7×7 convolution over the fused spectral feature $F_S$; it splits the feature map into $g_p$ groups of equal pixel count ordered by mask value. Feature-Grouped Attention projects each group into in-group and cross-group queries, keys, and a shared value, runs attention within a group and with the most similar other group chosen by cosine similarity, and combines the outputs with learnable weights. The spectral prompt makes the mask degradation-aware: Sobel captures high-frequency edges while SVD captures low-frequency structure, and linear attention lets the two cues inform each other before the mask is generated.","core_discovery":"SSGformer is an encoder-decoder transformer whose encoder is built from Spectral-based Decomposition Prompts, Mask Generators, and Spatial Grouping Transformer Blocks. The prompt runs the input through a Sobel operator and an SVD filter, refines both with separate blocks, fuses them with multi-head linear attention, and combines the results into a degradation-aware feature $F_S$. A single 7×7 convolution turns $F_S$ into a one-channel grouping mask $M_p$, which partitions the feature map at stage $p$ into $g_p$ equal-sized groups by mask value. Feature-Grouped Attention then applies in-group attention within each group and cross-group attention to the most similar group selected by cosine similarity over channel-pooled representatives, with learnable weights balancing the two streams; half the blocks attend along channels and half along space. The paper reports that this design reaches 32.63 dB PSNR on All-Weather and 24.77 dB on WeatherStream, with ablations showing that both the spectral prompt and the grouped attention contribute.","pith_inferences":["If the group selector's cosine-similarity choice is what drives the cross-group gain, then the same selector could be paired with any degradation-aware mask, including masks produced by learned prompt features; this would be a cheap test of whether the spectral prompt or the selection rule matters more.","The paper does not visualize the learned grouping masks or vary the group count $g_p$; measuring mask consistency across augmentations and reporting a group-count sensitivity curve would tell whether the mask is capturing degradation structure or just intensity thresholds.","Because Sobel and SVD are generic low-level cues, the same architecture could be pointed at other spatially varying corruptions such as motion blur, compression noise, or underwater scatter without retraining the spectral decomposer, though the paper only demonstrates weather."],"forward_implications":["On the All-Weather dataset, SSGformer reports 32.63 dB average PSNR and 0.939 average SSIM, the best overall scores, with per-task gains of +0.35 dB on rain and +0.18 dB on raindrop over the second-best method.","On the real-world WeatherStream dataset, SSGformer reports the best average PSNR at 24.77 dB, ahead of TransWeather, WGWS, and Histoformer.","Ablations show the full model beats variants with only Sobel or only SVD, only channel attention, and only in-group attention, so each proposed component contributes to the reported gain.","Because the model relies only on intra-image spectral and spatial cues, the approach is applicable without external knowledge sources such as LLMs or VLMs."],"supporting_citations":[{"why":"Supplies the All-Weather benchmark (Outdoor-rain, Snow100K-L, Raindrop) and the unified training protocol on which the central state-of-the-art comparison is made.","marker":"[28]"},{"why":"Histoformer is the strongest all-in-one baseline the method must beat and the source of the Pearson correlation loss and the pixel-sorting grouping idea it extends.","marker":"[47]"},{"why":"AdaIR represents the frequency-domain, global-filtering approach the paper argues is insufficient for localized weather degradations.","marker":"[13]"},{"why":"TransWeather is the all-in-one transformer baseline used in both All-Weather and WeatherStream comparisons.","marker":"[49]"},{"why":"WGWS supplies the weather-general/weather-specific baseline compared on both benchmarks and on real-world snow images.","marker":"[78]"},{"why":"WeatherStream provides the real-world rain, snow, and fog testbed used to evaluate cross-domain generalization.","marker":"[69]"},{"why":"Supplies the multi-head linear attention used inside the Spectral Feature Fusion Module to relate Sobel and SVD features.","marker":"[23]"},{"why":"Provides the RainDrop dataset for the raindrop sub-task in the All-Weather evaluation.","marker":"[37]"},{"why":"Provides the Outdoor-rain dataset for the rain sub-task in the All-Weather evaluation.","marker":"[27]"},{"why":"Provides the Snow100K-L dataset for the snow sub-task in the All-Weather evaluation.","marker":"[31]"}],"fun_headline_variants":["SSGformer: spectral grouping clears rain, snow, and haze","Grouped attention + spectral prompts hit 32.63 dB PSNR","One model, all weather: SSGformer beats all-in-one rivals","Sobel edges and SVD guide weather removal in SSGformer","Spectral-spatial grouping transformer tops 32.63 dB PSNR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single-convolution grouping mask partitions pixels into groups that actually track similar degradation, so that running attention inside and across those groups is more useful than ordinary attention; if the mask assignments are arbitrary or unstable, the reported gains would vanish.","fun_headline_variants_meta":{"raw":{"variants":["SSGformer: spectral grouping clears rain, snow, and haze","Grouped attention + spectral prompts hit 32.63 dB PSNR","One model, all weather: SSGformer beats all-in-one rivals","Sobel edges and SVD guide weather removal in SSGformer","Spectral-spatial grouping transformer tops 32.63 dB PSNR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1497,"prompt_tokens":976,"completion_tokens":521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":425}},"tokens_in":592,"tokens_out":521,"duration_ms":5291,"temperature":1.0,"reasoning_tokens":425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:35:59.607538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the learned grouping mask with a fixed random partition of pixels into $g_p$ equal groups and retrain; if All-Weather PSNR stays near 32.63 dB, the mask is not the source of the gain. A cheaper check is to compute the same SSGformer's mask on an image and its horizontally flipped or brightness-shifted version and measure overlap: low overlap would indicate the groups are not stable degradation regions.","supporting_citations":[{"cited_title":"All in one bad weather removal using architectural search","cited_arxiv_id":null,"evidence_quote":"Supplies the All-Weather benchmark (Outdoor-rain, Snow100K-L, Raindrop) and the unified training protocol on which the central state-of-the-art comparison is made."},{"cited_title":"Restoring images in adverse weather conditions via histogram transformer","cited_arxiv_id":null,"evidence_quote":"Histoformer is the strongest all-in-one baseline the method must beat and the source of the Pearson correlation loss and the pixel-sorting grouping idea it extends."},{"cited_title":"Transweather: Transformer-based restoration of images degraded by adverse weather conditions","cited_arxiv_id":null,"evidence_quote":"TransWeather is the all-in-one transformer baseline used in both All-Weather and WeatherStream comparisons."},{"cited_title":"Learning weather-general and weather-specific fea- tures for image restoration under multiple adverse weather conditions","cited_arxiv_id":null,"evidence_quote":"WGWS supplies the weather-general/weather-specific baseline compared on both benchmarks and on real-world snow images."},{"cited_title":"Weatherstream: Light transport au- tomation of single image deweathering","cited_arxiv_id":null,"evidence_quote":"WeatherStream provides the real-world rain, snow, and fog testbed used to evaluate cross-domain generalization."},{"cited_title":"Transformers are rnns: Fast autoregressive transformers with linear attention","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-head linear attention used inside the Spectral Feature Fusion Module to relate Sobel and SVD features."},{"cited_title":"Attentive generative adversarial network for raindrop removal from a single image","cited_arxiv_id":null,"evidence_quote":"Provides the RainDrop dataset for the raindrop sub-task in the All-Weather evaluation."},{"cited_title":"Heavy rain image restoration: Integrating physics model and conditional adversarial learning","cited_arxiv_id":null,"evidence_quote":"Provides the Outdoor-rain dataset for the rain sub-task in the All-Weather evaluation."},{"cited_title":"Desnownet: Context-aware deep net- work for snow removal","cited_arxiv_id":null,"evidence_quote":"Provides the Snow100K-L dataset for the snow sub-task in the All-Weather evaluation."}],"review_version":1}