{"id":"0f3d2174-c70c-4913-ba1d-4aeb7e4e273d","arxiv_id":"2411.12450","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A plug-and-play wavelet frequency guidance loss improves blind image restoration in diffusion models, giving up to 3.72 dB PSNR gain on motion deblurring.","lead":"This paper adds a frequency-domain correction to diffusion-based blind image restoration, using wavelet transforms to enforce detail consistency during sampling. The authors report PSNR gains up to 3.72 dB over the BlindDPS baseline on deblurring, with mixed metric results on ImageNet turbulence.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 14 is mathematically a reweighted version of the spatial L2 residual, so the 'frequency-domain consistency' claim needs a spatial high-pass baseline before it can be credited.","rationale":"The reader's weakest assumption concerned the lack of a spatial high-pass baseline and the heuristic nature of Eq. 14. My stress-test sharpens this into a concrete mathematical point: because the Haar DWT is an orthogonal transform, the high-frequency subband terms carry no information beyond the spatial residual; they merely reweight its frequency content. This does not refute the reported empirical gains, but it means the paper's central mechanistic claim is currently unsupported. A control experiment with an algebraically equivalent spatial high-pass loss would settle whether the frequency-aware framing is essential. Since the issue is addressable through a focused experiment and the reported gains are nontrivial, the existing CONDITIONAL verdict remains appropriate.","tokens_in":11933,"tokens_out":8640,"duration_ms":91726,"concrete_test":"Run the same BlindDPS pipeline with a control loss L_ctrl = ||d||^2 + lambda' * (||F_LH d||^2 + ||F_HL d||^2 + ||F_HH d||^2), where d = y - khat*xhat0 and F_LH, F_HL, F_HH are the same 2x2 Haar high-pass filters applied with stride 2 directly to the residual d, rather than to y and yhat separately. Tune lambda' on a held-out 100-image validation split and report mean and standard deviation over at least three sampling seeds on the FFHQ motion-blur set. Also verify analytically that DWT(y - yhat) equals DWT(y) - DWT(yhat), so this control is algebraically equivalent to Eq. 14. If L_ctrl reproduces Table 1's PSNR, FID, and LPIPS within noise, the DWT/frequency framing is not the causal ingredient and the paper should be reframed as a high-frequency reweighting study.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the wavelet subband terms in Eq. 14 add a genuinely new constraint beyond the spatial term ||y - khat*xhat0||^2. This premise is not supported by the mathematics. For a residual d = y - khat*xhat0, the four 2x2 Haar filters in Eq. 9 form an orthogonal (unnormalized) DWT; for each block, the energies of the LL, LH, HL, and HH subbands sum to a fixed multiple of the block's spatial energy. Consequently, Eq. 14 is, up to constants and boundary handling, a diagonal reweighting of the wavelet coefficients of that same spatial residual: it upweights high-frequency parts and downweights low-frequency parts. No information is added by 'frequency consistency' beyond the existing DPS data-consistency residual. The paper never ablates against a spatial-domain high-pass regularizer with matched total gradient scale, so the reported 3.72 dB gain cannot yet be attributed to the proposed frequency-aware mechanism. The empirical result may survive, but the central novelty claim would then reduce to 'reweight the existing data-consistency loss toward high frequencies,' not 'simultaneously enforce consistency in spatial and frequency domains.' This is load-bearing for the interpretation and significance of the contribution, though it does not by itself invalidate the empirical comparisons.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free, plug-and-play guidance loss for blind image restoration with diffusion models. The loss, defined in Eq. (14), combines the standard spatial data-consistency term of BlindDPS with additional terms that match high-frequency Haar wavelet subbands of the estimated degraded observation to those of the observed image. The authors evaluate the method on blind deblurring, imaging through turbulence, and multiple-degradation restoration, reporting consistent PSNR improvements over BlindDPS on FFHQ and, for deblurring and turbulence, a 3.72 dB gain on FFHQ motion deblurring. Ablation studies examine the effect of the weighting hyperparameter λ and of adding individual high-frequency subbands.","tokens_in":12211,"tokens_out":6335,"duration_ms":59270,"significance":"If the empirical gains hold, the paper offers a useful, simple, and training-free modification to BlindDPS that improves reconstruction fidelity in several blind restoration tasks. The main strengths are the clear experimental setup, the inclusion of ablations on λ and on frequency components, and the reporting of three metrics (PSNR, LPIPS, FID) across three tasks. However, the central novelty claim—that the loss simultaneously enforces consistency in the spatial and frequency domains—is undermined by the mathematics of the Haar DWT, which makes Eq. (14) a reweighted version of the spatial residual rather than a genuinely new constraint. The paper also contains a clear regression on ImageNet turbulence (FID and LPIPS worse than BlindDPS) that is understated in the text, and the 'various diffusion models' claim is not supported because only BlindDPS is used as the base method. These issues are fixable but require additional experiments and revised claims.","major_comments":[{"comment":"The mathematical claim that Eq. (14) enforces consistency 'in both the spatial and frequency domains' is not supported, because the Haar DWT used in Eqs. (9)–(13) is an orthogonal transform up to a constant scale. For the residual d = y − ŷ, the squared energies of the four subbands satisfy ||d_LL||² + ||d_LH||² + ||d_HL||² + ||d_HH||² = 4||d||². Substituting into Eq. (14) with a common weight λ gives L_freq = (1 + 4λ)||d||² − λ||d_LL||², i.e., a reweighting of the spatial residual that upweights high frequencies and downweights the LL component. No genuinely new data-consistency constraint is added beyond the DPS loss. The paper never compares against a spatial-domain high-pass regularizer of matched scale (e.g., ||∇(y − ŷ)||² or the residual minus its LL subband). Without that baseline, the reported gains in Tables 1 and 5 cannot be attributed to the proposed frequency-domain mechanism; they may simply reflect a beneficial reweighting of the existing data-consistency term. This is load-bearing for the paper's central novelty claim and should be addressed by adding the baseline or by revising the claim to 'high-frequency-weighted data consistency'.","section":"§3.3, Eq. (14)"},{"comment":"The ImageNet turbulence results do not support the general claim of improved perceptual quality. Compared with BlindDPS, the method improves PSNR (21.75 vs. 19.59) but worsens LPIPS (0.389 vs. 0.341) and FID (80.78 vs. 51.25). The text states that the method 'achieves comparable LPIPS results' on ImageNet, which understates a clear regression on two of three metrics. The paper should report and discuss this trade-off, and either temper the claim that the method provides 'the best visual quality' or explain why the ImageNet regression is acceptable (e.g., the domain gap from the FFHQ-pretrained diffusion model mentioned in the same section).","section":"§4.2, Table 2"},{"comment":"The claim that the guidance loss is 'plug-and-play' and 'can be integrated into various diffusion models' is not demonstrated: all experiments use the BlindDPS framework with its pretrained image and kernel diffusion models. To support the generality claim, the authors should either add an experiment with at least one different base method (e.g., DPS or IIGDM with a known forward operator) or revise the claim to 'can be integrated into BlindDPS.'","section":"§4"}],"minor_comments":[{"comment":"BlindDPS is cited as [7] in Tables 2 and 3, but the correct reference is [5]; reference [7] is the MCG paper. Please fix the citation inconsistencies throughout.","section":"Tables 2 and 3"},{"comment":"The notation y_i and ŷ_i is not defined; specify that these are the wavelet subbands of y and ŷ, respectively, and clarify whether a single λ is used for all three high-frequency subbands or a separate λ_i as written.","section":"§3.3, Eq. (14)"},{"comment":"The ablation on λ is performed on only 100 images and without error bars; given the stochastic nature of diffusion sampling, report variance or multiple seeds.","section":"§4.4, Table 4"},{"comment":"'Two reserve diffusion processes' is a typo for 'two reverse diffusion processes.'","section":"§2.1"},{"comment":"The phrase 'simultaneously enforces content consistency in both the spatial and frequency domains' overstates the mechanism in light of the orthogonality argument; consider rewording to describe the loss as a frequency-weighted spatial consistency term.","section":"§1 and §5"}],"recommendation":"major_revision","confidential_remarks":"The paper would be considerably strengthened by a direct comparison against a spatial high-pass baseline; without it, the contribution may be viewed as an incremental reweighting of the BlindDPS loss. The ImageNet regression and the citation errors should be fixed before acceptance. I do not see circularity or misconduct; the main issue is overclaiming the mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is thinner than it looks, but the empirical result is probably real. The proposed guidance loss adds a Haar-wavelet high-frequency term to the BlindDPS data-consistency loss and reports solid PSNR gains on blind deblurring, turbulence, and multiple degradations. On FFHQ motion deblurring the 3.72 dB gain over BlindDPS is the kind of number that gets attention; the visual examples line up with it.\n\nThe problem is the interpretation. Because the Haar wavelets in Eq. 9 form an orthogonal (up to scale) transform, the squared difference between the degraded estimates and observations in the three high-frequency subbands, plus the spatial L2 term, is equivalent to a single spatial L2 loss with reweighted frequencies: the LL band gets weight 1, the HF bands get 1+λ. No new 'frequency consistency' constraint is added. The paper never ablates against a spatial high-pass residual of matched gradient scale, so the reported gains are plausibly just from upweighting high-frequency errors, not from a genuinely dual-domain mechanism.\n\nOther soft spots: no error bars or repeated runs; the ImageNet turbulence results in Table 2 show FID going from 51.25 to 80.78 and LPIPS from 0.341 to 0.389 while the text calls the LPIPS 'comparable'—that's a regression they walk past. The hyperparameter λ is chosen on the evaluation set (they sweep on 100 FFHQ images and keep 0.1), so the headline numbers are upper-bound-y. They also don't mention code availability.\n\nCredit where due: the method is simple, training-free, plug-and-play, and the ablations over λ and over which HF subbands to include are informative. The paper is honest about the FFHQ-pretrained prior causing a domain gap on ImageNet. The citation pattern is fine; the closest related work on wavelet losses is in trained restoration networks, and the diffusion-specific combination is not already in the literature.\n\nBottom line: this is a useful empirical note whose central conceptual claim does not survive contact with Parseval's theorem. A serious referee should not reject it—the fix is an experimental section with a spatial high-pass baseline and honest framing—but the paper needs major revision. It is worth citing downstream as a weighting trick, not as evidence for dual-domain consistency.","headline":"Eq. 14 is a frequency-weighted spatial residual, not a new dual-domain constraint, but the simple plug-in still gets real PSNR gains and deserves a careful revision.","tokens_in":12726,"tokens_out":3224,"would_cite":false,"duration_ms":32915,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A frequency-aware guidance loss based on the 2D Haar wavelet transform, added to diffusion-based blind image restoration, enforces consistency in both spatial and high-frequency domains and improves motion-deblurring PSNR by 3.72 dB over…","keywords":["blind image restoration","diffusion models","frequency-aware guidance","discrete wavelet transform","posterior sampling","image deblurring","turbulence restoration","zero-shot restoration"],"falsifier":"Compare the proposed loss against a purely spatial high-pass baseline, such as an $\\ell^2$ penalty on the Laplacian or gradient of $\\hat{y}$ versus $y$, on the same FFHQ deblurring benchmark. If the gains over BlindDPS disappear, the improvement is due to edge emphasis rather than wavelet-specific frequency consistency.","tokens_in":11754,"feed_emoji":"🖼️","tokens_out":4289,"duration_ms":36418,"temperature":0.7,"pith_summary":"This paper proposes a training-free guidance loss for diffusion-based blind image restoration. The loss adds a frequency-domain term, computed with a two-dimensional Haar discrete wavelet transform, to the usual spatial consistency term used in posterior sampling. Plugged into the BlindDPS framework, it enforces consistency between the estimated degraded observation and the observed image in both spatial and high-frequency wavelet subbands. The authors report a 3.72 dB PSNR improvement over BlindDPS for motion deblurring on FFHQ, along with consistent gains in imaging through turbulence and mixed blur-plus-noise restoration.","feed_headline":"Wavelet guidance lifts blind deblurring by 3.72 dB","feed_subtitle":"A plug-in loss enforcing wavelet-domain consistency during sampling sharpens blind deblurring, turbulence, and mixed-degradation results.","key_machinery":"The load-bearing mechanism is the two-dimensional discrete wavelet transform with the four orthogonal Haar filters $k_{LL}$, $k_{LH}$, $k_{HL}$, and $k_{HH}$ of Eq. (9), applied with stride 2 to the estimated degraded observation $\\hat{y}$. The low-frequency subband carries the spatial consistency term, while the three high-frequency subbands carry the new frequency guidance. At each sampling step, the gradient of the combined loss is backpropagated through $\\hat{x}_0$ and $\\hat{k}_0$, redirecting the reverse diffusion so that reconstructed details match the observed image's high-frequency structure.","core_discovery":"The central claim is that existing diffusion guidance, which measures data consistency only in the spatial domain, leaves high-frequency content unconstrained and thus produces distorted, texture-poor results. The paper argues that adding a weighted penalty on the high-frequency subbands ($LH$, $HL$, $HH$) of the two-dimensional Haar wavelet decomposition of $\\hat{y} = \\hat{k}_0 * \\hat{x}_0$ versus those of the observation $y$ steers the sampling trajectory toward cleaner, more detailed images. This claim is embodied in the loss $\\mathcal{L}_{\\text{freq}} = \\|y - \\hat{y}\\|_2^2 + \\sum_{i \\in \\{LH,HL,HH\\}} \\lambda_i \\|\\hat{y}_i - y_i\\|_2^2$, which can be added plug-and-play to pre-trained diffusion models and requires no training.","pith_inferences":["The frequency penalty is only as reliable as the estimated kernel: if $\\hat{k}_0$ is wrong, matching $\\hat{y}$'s high-frequency subbands to $y$'s could push content in the wrong direction. A direct test would compare frequency guidance computed on $\\hat{x}_0$ versus $\\hat{y}$.","Treating the high-frequency penalty as a sharpness prior suggests it may transfer to other inverse problems where texture fidelity matters, such as super-resolution, without architectural change.","Ablating the wavelet-specific choice against a simple spatial high-pass filter (for example, an $\\ell^2$ penalty on the Laplacian or gradient of $\\hat{y}$ versus $y$) would clarify whether the benefit comes from the wavelet subbands or just from emphasizing edges."],"forward_implications":["If the claim holds, any guided diffusion restoration method with a differentiable forward model can adopt this frequency loss without retraining.","The weight $\\lambda$ gives a direct handle on the reconstruction-perception tradeoff: larger values emphasize high-frequency fidelity at the risk of distortion, while smaller values produce smoother results.","Because the loss operates on the estimated degraded observation, it remains applicable when the degradation kernel is unknown and estimated jointly, as in BlindDPS.","The reported gains in mixed blur-plus-noise restoration suggest the frequency penalty is robust to more than one degradation at a time."],"supporting_citations":[{"why":"BlindDPS is the baseline and the framework this method extends; it supplies the joint estimation of clean image and degradation kernel.","marker":"[5]"},{"why":"DPS provides the posterior-sampling gradient approximation that the spatial consistency term is built on.","marker":"[6]"},{"why":"DDPM defines the forward and reverse diffusion processes and the Tweedie-style estimate of $\\hat{x}_0$ used in the guidance.","marker":"[15]"},{"why":"Prior work using wavelet transforms in image restoration, which the paper contrasts with its training-free per-step frequency guidance.","marker":"[23]"},{"why":"Earlier frequency-selection work in restoration, cited as related wavelet-based frequency learning that the method extends.","marker":"[9]"},{"why":"Provides evidence that high-frequency components drive perceptual quality, justifying the added high-frequency regularization.","marker":"[10]"}],"fun_headline_variants":["Frequency-aware guidance boosts blind deblurring by 3.72 dB","Wavelet loss sharpens blind restoration, adds 3.72 dB","Plug-and-play frequency guidance lifts diffusion deblurring","High-frequency guidance cuts distortion in blind restoration","Spatial + frequency consistency: +3.72 dB for blind deblurring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that matching the high-frequency wavelet subbands of the predicted degraded image $\\hat{y}$ to those of the observed image $y$ is a valid signal for recovering the clean image, even though this penalty is a heuristic add-on rather than a consequence of the posterior.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-aware guidance boosts blind deblurring by 3.72 dB","Wavelet loss sharpens blind restoration, adds 3.72 dB","Plug-and-play frequency guidance lifts diffusion deblurring","High-frequency guidance cuts distortion in blind restoration","Spatial + frequency consistency: +3.72 dB for blind deblurring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000599,"raw_usage":{"total_tokens":2787,"prompt_tokens":922,"completion_tokens":1865,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1773}},"tokens_in":538,"tokens_out":1865,"duration_ms":11056,"temperature":1.0,"reasoning_tokens":1773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:30:34.777977+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the proposed loss against a purely spatial high-pass baseline, such as an $\\ell^2$ penalty on the Laplacian or gradient of $\\hat{y}$ versus $y$, on the same FFHQ deblurring benchmark. If the gains over BlindDPS disappear, the improvement is due to edge emphasis rather than wavelet-specific frequency consistency.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"BlindDPS is the baseline and the framework this method extends; it supplies the joint estimation of clean image and degradation kernel."},{"cited_title":"ICLR (2022)","cited_arxiv_id":null,"evidence_quote":"DPS provides the posterior-sampling gradient approximation that the spatial consistency term is built on."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops","cited_arxiv_id":null,"evidence_quote":"Prior work using wavelet transforms in image restoration, which the paper contrasts with its training-free per-step frequency guidance."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)","cited_arxiv_id":null,"evidence_quote":"Earlier frequency-selection work in restoration, cited as related wavelet-based frequency learning that the method extends."},{"cited_title":"In: Proceed- ings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Provides evidence that high-frequency components drive perceptual quality, justifying the added high-frequency regularization."}],"review_version":1}