{"id":"45d7947e-a516-414f-8d39-21d325ffbf68","arxiv_id":"2505.05504","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SWFormer is a multi-domain image restoration backbone that combines spatial, wavelet, and Fourier processing and achieves competitive state-of-the-art results across ten tasks with lower computational cost.","lead":"This paper introduces SWFormer, an image restoration network that combines spatial, wavelet, and Fourier domain processing inside a transformer-like architecture. It reports strong results across ten restoration tasks while using fewer parameters and lower latency than many prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is unqualified: only SWFormer-l wins most benchmarks, while the small/medium variants highlighted for efficiency underperform published baselines on many datasets.","rationale":"The reader's weakest assumption (multi-domain prior) is not where the claim is most vulnerable: the network is trainable and could still work even if the prior is not universal, so it mainly affects the explanation rather than the SOTA assertion. The reviewer's own rationale already notes the overclaim for smaller variants; I make that the primary concern. This is a correctness-of-claim issue, not an internal inconsistency, and it is fixable by qualification and a Pareto analysis. I also note that the 'lossless' property of the learnable wavelet transform (Sec. III-B2) needs a constraint or explicit justification, but it is secondary to the variant-level overclaim. The conditional verdict remains appropriate, so no change to the reader's verdict is needed.","tokens_in":26685,"tokens_out":14744,"duration_ms":151041,"concrete_test":"Build a per-variant win/loss matrix from Tabs. I-X: for each benchmark and metric, mark whether SWFormer-s/m/l beats the best published baseline (PSNR/SSIM, and LPIPS where reported). Then test the abstract's claim against each variant: if SWFormer-s wins on fewer than half the PSNR benchmarks, replace the unqualified 'outperforms SOTA' with a variant-specific statement and add a Pareto plot of PSNR vs latency/FLOPs showing which variant, if any, is on the efficiency frontier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('outperforms state-of-the-art methods', 'SOTA on most datasets') is stated without specifying which SWFormer variant supports it. The efficient variants used to motivate the trade-off do not actually achieve SOTA. From Tabs. II-VIII: on DDN-Data SWFormer-s scores 34.18 dB vs NeRD-Rain 34.45; on DID-Data 35.18 vs FADformer 35.48; on SPA-Data 49.03 vs SFHformer 50.11; on all five dehazing sets SWFormer-s is below SFHformer (e.g., ITS 41.86 vs 43.03, O-HAZE 22.90 vs 25.81); on GoPro 33.55 vs MISC Filter 34.10 and on HIDE 31.27 vs X-Restormer 31.76; on L-400 30.16 vs SFHformer 30.18; on CSD 36.85 vs SFHformer 37.45. Even the largest variant loses PSNR on DPDD (26.220 vs GGKMNet 26.272). Thus the abstract's unqualified SOTA claim is only defensible for SWFormer-l on 'most datasets', while the efficiency headline in Fig.1(b)/Tab.XI uses SWFormer-s. No single configuration is demonstrated to be both SOTA and maximally efficient, so the 'excellent balance' claim needs a per-variant qualification or a Pareto analysis before it can be accepted as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SWFormer, a Transformer-like image-restoration backbone that replaces self-attention with a Spatial-Wavelet-Fourier Mixer (SWFM) and uses a Multi-scale ConvFFN (MSFN), all embedded in a Lossless Multi-Input Multi-Output (LMIMO) framework. The authors motivate the design with the qualitative observation that degradation energy concentrates in specific wavelet sub-bands and Fourier frequencies, allowing local-region-global receptive field modeling. They evaluate the method on 26 benchmarks across 10 restoration tasks, comparing three model variants (small, medium, large) against prior state-of-the-art methods in terms of PSNR/SSIM (and sometimes LPIPS), parameter count, FLOPs/MACs, and latency. The central claims are that SWFormer outperforms state-of-the-art methods on most datasets and achieves a favorable efficiency-performance trade-off.","tokens_in":27050,"tokens_out":5518,"duration_ms":58701,"significance":"If the claims hold, SWFormer-l would be a practical and reasonably efficient multi-task restoration backbone: it obtains the best or near-best PSNR on most of the 26 benchmarks, with a large reduction in parameters and FLOPs relative to several transformer baselines. The paper's strengths include the breadth of the experimental study, the inclusion of complexity and latency measurements, ablations of the main components, and a public code release. The performance margins over prior work are often small (0.1 to 0.4 dB), but they are consistent for the large variant on most datasets. A notable weakness is that the unqualified SOTA claim is not supported by the small and medium variants, and the \"lossless\" property of the LMIMO framework is not formally established. The empirical evaluation is conducted on external benchmarks, so there is no circularity concern from fitting-then-testing on the same data.","major_comments":[{"comment":"The abstract and Section I claim that the proposed model \"outperforms state-of-the-art methods\" and achieves SOTA \"on most datasets\" without specifying which SWFormer variant supports this claim. The tables show that only SWFormer-l supports \"most datasets,\" while SWFormer-s and SWFormer-m, which are the variants highlighted for efficiency in Fig. 1(b) and Table XI, are not SOTA. For example, in Table II SWFormer-s is below NeRD-Rain on DDN-Data (34.18 vs. 34.45 dB) and below SFHformer on SPA-Data (49.03 vs. 50.11 dB); in Table IV SWFormer-s is below SFHformer on all five dehazing benchmarks (ITS 41.86 vs. 43.03 dB, O-HAZE 22.90 vs. 25.81 dB); in Table VII SWFormer-s is below SFHformer on L-400 (30.16 vs. 30.18 dB). Since the efficiency claims are based on SWFormer-s/m, no single configuration is demonstrated to be simultaneously SOTA and the most efficient. The SOTA claim should be qualified per variant, or the paper should provide a Pareto analysis of the s/m/l family. This is a load-bearing issue for the central claim.","section":"Abstract and Section I, Tables II, IV, V, VII"},{"comment":"The term \"Lossless\" in Lossless Multi-Input Multi-Output is not established. The authors state that the multi-input part uses Wavelet Transform or Pixel Unshuffle for downsampling, \"which preserves lossless information,\" but no proof or constraint is given. A wavelet decomposition is lossless only for orthogonal or perfect-reconstruction filter banks. In the wavelet branch, the transform W_phi is described as a \"learnable 2D wavelet transform\" with parameterized convolutions, and W^{-1} is a separate learnable inverse, with no orthogonality or perfect-reconstruction constraint described. If a fixed orthogonal wavelet (e.g., Haar) is used in the LMIMO input, this should be stated explicitly; if the transform is learnable, the paper should either impose a perfect-reconstruction constraint or provide an empirical reconstruction-error measurement. As written, \"lossless\" is an unsupported naming claim.","section":"Section III-A, LMIMO"},{"comment":"The Fourier-domain branch as written is not mathematically well-defined. The text says that after applying the 2D FFT, the real and imaginary parts are concatenated to produce real tensors F1_j and F2_j in R^{H x W x 2C}, and Eq. (5) then applies the inverse FFT to their element-wise product. The inverse FFT requires a complex input. For Eq. (5) to be well-posed, the authors must state that the 2C channels are split into real and imaginary halves and that the gating mechanism operates on each half, or they should formulate the branch using C-channel complex tensors. Without this clarification, the core Fourier branch cannot be reproduced from the description.","section":"Section III-B, Eq. (5)"}],"minor_comments":[{"comment":"The caption of Table XII reads \"Experimental Results on Model Complexity,\" but the table reports inference latency and PSNR/SSIM; the caption should be renamed to \"Inference Latency\" or similar.","section":"Section IV-L, Table XII"},{"comment":"The notation {I_i, i = 1, 2, 3} in {R^{H/n x W/n x 3 x 2^{n-1}}, n = 1, 2, 3} is confusing: the expression 3 x 2^{n-1} appears to refer to the number of channels after wavelet decomposition, but writing it as a set membership over n is not standard and should be clarified with an explicit per-scale channel mapping.","section":"Section III-A"},{"comment":"The multi-domain loss in Eq. (6) introduces a hyperparameter lambda set to 0.1, but no ablation or sensitivity study for lambda is reported; the authors should at least state whether the results are stable with respect to this choice.","section":"Section IV-M, Eq. (6)"},{"comment":"There are several typos and minor language issues: \"As reflected by the its name\" in Section III-B, \"we replaces self-attention\" in the Conclusion, \"themself\" in Section II-A, and a stray space in the Section IV-F heading \"F . Results on Defocus Deblurring Task.\" These should be corrected in a final revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is an extension of the authors' own ECCV 2024 work (SFHformer), and the main comparisons are against that prior model. The authors should clarify in the final version whether all baseline numbers are taken from the original papers or re-run under identical settings; this is important because the reported margins over some baselines are small and protocol differences in resolution, training schedule, or evaluation code could change the ranking on individual benchmarks. The scope is broad for a single paper, and the incremental contribution over the conference version should be clearly delineated in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth taking seriously, but the abstract oversells it. The real contribution is a Transformer-like restoration backbone that replaces self-attention with a spatial-wavelet-Fourier token mixer, plus a multi-scale FFN and an LMIMO inter-block design that produces small/medium/large variants from one trained model. That combination is new, and the evaluation is unusually broad: 26 benchmarks across 10 tasks, with ablations and latency numbers. They also compare directly against their own SFHformer (ECCV 2024), which is the right baseline and keeps the incremental contribution visible.\n\nWhat it does well: the efficiency story is credible. SWFormer-s is 2.5x faster than SFHformer on Rain200L and beats it there, and the Pareto-style comparison in Tab. XI shows the trade-off clearly. The Fourier gating versus dynamic convolution ablation (Tab. XV) is a good example of an honest latency/performance trade-off. The wavelet-branch idea is plausible and the visual motivation in Fig. 1(c) is suggestive.\n\nThe soft spots are real and mostly about claims, not the method. The stress-test note is correct: the unqualified \"outperforms SOTA\" in the abstract is only defensible for SWFormer-l. SWFormer-s loses to SFHformer on every dehazing set (e.g., ITS 41.86 vs 43.03, O-HAZE 22.90 vs 25.81), and also on DDN, DID, SPA, GoPro, HIDE, L-400, and CSD. The small and medium variants are efficient and competitive, but they are not SOTA. The paper needs a per-variant qualification or a Pareto analysis before that headline is honest.\n\nSecond, \"Lossless Multi-Input Multi-Output\" overclaims. Pixel Unshuffle is lossless; a learnable wavelet convolution is not automatically a perfect reconstruction filter bank. If they use a fixed orthogonal wavelet, fine, but they say \"learnable\" and parametrize it with convolutions. They need to either constrain it to be an orthogonal filter bank or show reconstruction error. As written, the name promises something the paper doesn't establish.\n\nMinor: the qualitative prior in Fig. 1(c) is not quantified across tasks, and the loss weight lambda=0.1 has no sensitivity analysis. Both are fixable in revision.\n\nOverall, the architecture is sound and the experiments are extensive enough to justify referee time. The reader's conditional verdict is fair. The paper needs a revised abstract and a careful pass on the lossless claim. I would bring it to a reading group focused on efficient restoration backbones, and I would send it to peer review, expecting revision rather than acceptance as is.","headline":"SWFormer is a genuinely efficient and broadly tested restoration backbone, but the SOTA claim only holds for the largest variant and the 'lossless' wavelet claim needs proof before the paper is publishable as written.","tokens_in":27539,"tokens_out":1842,"would_cite":false,"duration_ms":20790,"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":"One network handles ten image-restoration tasks by mixing spatial, wavelet, and Fourier features.","keywords":["image restoration","multi-domain learning","wavelet transform","Fourier transform","transformer","deraining","dehazing","low-light enhancement"],"falsifier":"For each of the ten tasks, compute the fraction of degradation energy in each wavelet sub-band and Fourier band; if any task (for example, shadow removal or cloud removal) shows degradation spread roughly uniformly across sub-bands, the central prior is violated. A direct behavioral test is the paper's own swap experiment: swap the identified sub-bands between clear and degraded pairs and measure PSNR, and if for some task the swapped clear image is not close to the degraded image, the prior fails there.","tokens_in":1614,"feed_emoji":"","tokens_out":3456,"duration_ms":81231,"temperature":0.7,"pith_summary":"The paper tries to establish that a single Transformer-style backbone, SWFormer, can outperform task-specific state-of-the-art models across ten different image restoration tasks by exploiting a prior shared by all degradations: in Fourier space and in wavelet sub-bands, the difference between a degraded image and a clean one is concentrated rather than spread out. It supports this with qualitative swap experiments and with a redesigned token mixer that replaces self-attention with spatial, wavelet, and Fourier branches, together with a multi-scale feed-forward network and a multi-input/multi-output encoder-decoder. The payoff, if correct, is practical: high restoration quality without the heavy training and inference cost that self-attention-based restoration models carry. The paper reports state-of-the-art results on most of twenty-six benchmarks while using a fraction of the parameters, FLOPs, and latency of comparable transformers.","feed_headline":"One backbone tops most of ten image-restoration tasks","feed_subtitle":"Swap self-attention for spatial-wavelet-Fourier mixing: lower latency, higher PSNR on deraining, deblurring, and dehazing.","key_machinery":"The central object is the Spatial-Wavelet-Fourier Mixer (SWFM), the token mixer that replaces vanilla self-attention. It splits an expanded feature into three branches: a depthwise convolution for local pixel-level features, a learnable two-dimensional wavelet transform followed by depthwise convolution and inverse wavelet transform for region-level sub-band features, and an FFT-based branch with a gating mechanism for global frequency filtering. The three branches are fused through channel attention, giving local-region-global receptive-field modeling at O(N log N) cost instead of the O($N^{2}$) cost of self-attention. The companion modules are the Multi-scale ConvFFN (MSFN), which aggregates multi-resolution features from those domains, and the Lossless Multi-Input Multi-Output (LMIMO) architecture, which feeds wavelet- or pixel-unshuffle-downsampled inputs and reconstructs outputs at several scales.","core_discovery":"The paper's central claim is that looking at image degradation through a spatial-wavelet-Fourier lens reveals a common prior: in the Fourier domain, degradations such as rain, blur, haze, and low light are compact rather than spatially diffuse, and after a wavelet transform the degradation concentrates in specific sub-bands (for example, LL and HL for rain, LL for haze and low light). Swapping those sub-bands between a clear and a degraded image swaps the degradation itself, which the paper takes as evidence that the domain separation is a real structural property. On that basis SWFormer replaces the self-attention token mixer with a three-branch Spatial-Wavelet-Fourier Mixer that runs depthwise convolution, a learnable wavelet transform, and an FFT/IFFT gating branch in parallel, and replaces the usual feed-forward network with a multi-scale ConvFFN. Inside a lossless multi-input/multi-output encoder-decoder, the same trained model can produce small, medium, and large outputs. The paper reports that SWFormer reaches state-of-the-art PSNR/SSIM on most of twenty-six benchmarks across ten restoration tasks while using fewer parameters, fewer FLOPs, and lower latency than the transformers it compares against.","pith_inferences":["The paper leaves the sub-band-separation claim qualitative; a direct next step is to quantify the fraction of degradation energy in each wavelet sub-band and Fourier band per task, and to verify that tasks with weaker separation lose less from removing that branch.","Because the wavelet branch uses learnable convolutions, one could test whether the learned filters converge to standard wavelet filters and whether the selected sub-bands shift across tasks, turning the prior into a measurable design rule.","The success across ten degradations suggests the multi-domain mixer could be transferred to neighboring low-level problems such as JPEG artifact removal or video restoration, but that extrapolation goes beyond the evidence in this paper."],"forward_implications":["Training a single SWFormer configuration on one task can serve as a strong shared encoder or initializer for other restoration tasks, because the multi-domain prior is treated as task-agnostic.","For deployment, SWFormer-s offers a low-latency option: at 256x256 resolution it restores Rain200L in 25.9 ms while beating the previous best deraining method in PSNR, and at 1024x1024 resolution it is much faster than DRSformer with higher PSNR.","The LMIMO design yields small, medium, and large variants from one training run, so a user can choose a compute budget without retraining separate models.","The Fourier branch's O(N log N) complexity makes the architecture suitable for high-resolution inputs where quadratic self-attention becomes prohibitive.","If the sub-band-separation prior holds, the same architectural recipe generalizes to degradations beyond the ten tested tasks, such as combining multiple degradations in one input image."],"supporting_citations":[{"why":"Introduces the self-attention Transformer for vision, the mechanism that SWFM is designed to replace.","marker":"[24]"},{"why":"Defines the MetaFormer-style token-mixer/FFN block structure that SWFormer keeps and modifies.","marker":"[101]"},{"why":"NAFNet is the simple baseline whose overall design and block structure inspire SWFormer.","marker":"[8]"},{"why":"The authors' conference version SFHformer is the direct predecessor whose Fourier dynamic convolution is replaced by a gating filter.","marker":"[43]"},{"why":"Restormer supplies the channel self-attention and gated FFN baselines that SWFormer compares against in most tasks.","marker":"[103]"},{"why":"DRSformer is the sparse-attention deraining baseline used for the main efficiency and latency comparison.","marker":"[12]"},{"why":"Uformer supplies the window-attention token mixer baseline and the DFFN baseline used in the ablation study.","marker":"[92]"}],"fun_headline_variants":["Spatial-wavelet-Fourier mixer replaces self-attention for restoration","SWFormer: one model, ten restoration tasks, low latency","Multi-domain token mixer restores images across ten tasks","SWFormer: spatial-wavelet-Fourier mixer for restoration","From haze to rain: SWFormer handles all ten degradations"],"cache_read_input_tokens":29696,"weakest_assumption_plain":"The whole design assumes that across all ten degradations, the difference between a degraded image and a clear image stays concentrated in a few wavelet sub-bands and Fourier frequencies, so that processing those domains separately is enough; if that concentration fails for a given degradation, the multi-branch mixer could lose to simpler efficient backbones.","fun_headline_variants_meta":{"raw":{"variants":["Spatial-wavelet-Fourier mixer replaces self-attention for restoration","SWFormer: one model, ten restoration tasks, low latency","Multi-domain token mixer restores images across ten tasks","SWFormer: spatial-wavelet-Fourier mixer for restoration","From haze to rain: SWFormer handles all ten degradations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001312,"raw_usage":{"total_tokens":5400,"prompt_tokens":1051,"completion_tokens":4349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":4262}},"tokens_in":667,"tokens_out":4349,"duration_ms":28987,"temperature":1.0,"reasoning_tokens":4262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:36:31.118623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For each of the ten tasks, compute the fraction of degradation energy in each wavelet sub-band and Fourier band; if any task (for example, shadow removal or cloud removal) shows degradation spread roughly uniformly across sub-bands, the central prior is violated. A direct behavioral test is the paper's own swap experiment: swap the identified sub-bands between clear and degraded pairs and measure PSNR, and if for some task the swapped clear image is not close to the degraded image, the prior fails there.","supporting_citations":[{"cited_title":"In: IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Defines the MetaFormer-style token-mixer/FFN block structure that SWFormer keeps and modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The authors' conference version SFHformer is the direct predecessor whose Fourier dynamic convolution is replaced by a gating filter."},{"cited_title":"In: IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Restormer supplies the channel self-attention and gated FFN baselines that SWFormer compares against in most tasks."},{"cited_title":"In: IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Uformer supplies the window-attention token mixer baseline and the DFFN baseline used in the ablation study."}],"review_version":1}