{"id":"52cff200-d41a-46ec-a532-d9912abfe41d","arxiv_id":"2507.01275","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FrDiff performs unpaired image dehazing by diffusing over amplitude residuals between hazy and clear images, with a phase correction module, and reports state-of-the-art results on RESIDE-based benchmarks.","lead":"A new dehazing method, FrDiff, trains a diffusion model to reconstruct the amplitude spectrum of clear images and uses it to remove haze from unpaired hazy photos. It reports higher PSNR and SSIM than prior unpaired dehazing methods on synthetic benchmarks and better no-reference scores on real-world sets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own supplementary contradicts its headline SOTS-Indoor number (Tab. A2: 36.54 dB vs Tab. 1: 27.43 dB for λ=1), so the central empirical claim is not internally reproducible.","rationale":"The central claim of the paper is quantitative: FrDiff's 27.43 dB PSNR and 0.957 SSIM on SOTS-Indoor make it the best unpaired dehazing method in Tab. 1. The load-bearing condition is not the architecture's plausibility but that the number is produced by the described pipeline under the described evaluation. Supp. Tab. A2 violates that condition: it gives 36.54 dB for the default hyperparameter setting on the same dataset. I checked whether this could be a harmless formatting slip: the sweep values are all near 36.5 dB, not near 27.4 dB, and the table is explicitly described as 'Results of different λ... on SOTS-Indoor.' Even a single bad row would be a problem, but the entire table disagrees with Tab. 1. The manuscript gives no protocol note that would reconcile the numbers. The reader's stated weakest assumption was the global-amplitude premise; I agree with that as a scientific limitation (and the authors concede it in Supp. A7), but it is not the most load-bearing threat. A method whose frequency assumption is imperfect can still support its claim if the measurement is credible; a method whose own reported numbers disagree does not. Since the reader already made the verdict conditional on resolving this discrepancy and releasing code/weights, my stress-test does not move the verdict; it sharpens the reason. I recommend UNCHANGED, i.e., CONDITIONAL until one SOTS-Indoor number is reproduced with released code.","tokens_in":19832,"tokens_out":7747,"duration_ms":95899,"concrete_test":"Release the trained FrDiff weights and the exact evaluation script, then rerun the SOTS-Indoor evaluation under the default λ=1. If the obtained PSNR is ≈36.5 dB, Tab. 1's 27.43 dB is wrong and the SOTA comparison must be re-drawn; if it is ≈27.4 dB, Supp. Tab. A2 is wrong; if it matches neither, the evaluation protocol is inconsistent. The discrepancy is settled only when one of the two published numbers is reproduced exactly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Supp. Tab. A2 reports FrDiff on SOTS-Indoor at 36.54 dB PSNR when λGAN=λPatchNCE=λdiff=1, the exact default setting used for the model in the main paper. Main Tab. 1 reports 27.43 dB PSNR for the same model and dataset. The sweep in Tab. A2 is flat (36.31–36.58 dB), so hyperparameter sensitivity cannot explain a 9.11 dB gap. No statement in the text indicates a different split, resolution, or metric implementation for Tab. A2. The central claim that FrDiff outperforms state-of-the-art unpaired dehazing is therefore carried by numbers that the manuscript itself does not consistently reproduce. If Tab. A2 is the correct evaluation, the reported advantage is understated by roughly 9 dB; if Tab. 1 is correct, Tab. A2 is an unexplained artifact. Either way, the empirical foundation for the headline claim is not settled. This is more immediate than the spatially-varying-haze limitation acknowledged in Supp. A7: a conceptually plausible method still needs a trustworthy measurement, and the paper currently lacks one because no code, weights, or seeds are provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FrDiff, an unpaired image dehazing method built around a frequency-domain diffusion model. The method first computes an amplitude residual z by aligning the global mean and standard deviation of the hazy amplitude spectrum to that of an unpaired clear image (Eq. 3). This residual is used both as a training target and as input to a Frequency Compensation Layer in a NAFNet-based dehazing network, while a Phase Correction Module refines the phase spectrum. Training is split into two stages: stage one trains the dehazing network with the ARE residual directly; stage two jointly trains a small denoising network and the dehazing network with an L1 loss on the final reverse-diffusion sample (Eq. 11). Experiments on SOTS-Indoor/Outdoor, HSTS-Synth, I-HAZE, O-HAZE, and real-world sets are reported, with the main table claiming 27.43 dB PSNR on SOTS-Indoor.","tokens_in":20159,"tokens_out":9016,"duration_ms":97238,"significance":"If the reported numbers are reliable, the paper would be a meaningful contribution to unpaired dehazing: it introduces a frequency-domain perspective, shows that a lightweight diffusion model (8.76M parameters, 54.56 GFLOPs at T=8) can be conditioned on amplitude spectra, and provides detailed ablations, algorithm pseudocode, and supplementary dark-channel statistics supporting the amplitude-based motivation. The central claim, however, is currently not internally reproducible. The main table reports 27.43 dB on SOTS-Indoor, while the supplementary hyperparameter table reports 36.54 dB for the same model and default settings, a 9.11 dB inconsistency. The absence of code, weights, seeds, and error bars further prevents verification. The conceptual idea is plausible and the paper is generally clearly written, but the empirical foundation of the headline result needs to be resolved before the contribution can be accepted.","major_comments":[{"comment":"The main paper reports 27.43 dB PSNR for FrDiff on SOTS-Indoor (Table 1; also the 'Full model' row of Table 3), while Supplementary Table A2 reports 36.54 dB for the same model with the same default hyperparameters (lambda_GAN = lambda_PatchNCE = lambda_diff = 1). The sweep in Table A2 is essentially flat (36.31-36.58 dB), so hyperparameter sensitivity cannot explain a 9.11 dB gap, and no text statement indicates a different split, resolution, or metric implementation for the two tables. The manuscript therefore contains two mutually inconsistent measurements of its central result, and the headline claim that FrDiff outperforms prior unpaired dehazing methods is not internally reproducible. Please correct the discrepancy, specify exactly which protocol produced each table, and provide the code or weights needed to verify the reported numbers.","section":"Table 1 vs. Supplementary Table A2"},{"comment":"The 'w/o DM' ablation uses the ARE residual directly, and the full model uses a DM that is trained with L_diff = ||z - z_hat||_1 (Eq. 11) to reproduce that same residual. The 2.77 dB gain of the full model over 'w/o DM' therefore conflates the effect of the DM architecture with the two-stage joint-training protocol; the 'w/o JT' row already shows a 0.66 dB gain from joint training alone. Please clarify what the DM contributes beyond a learned regressor from the hazy amplitude spectrum to the ARE-defined target, and provide an ablation in which the DM is replaced by a deterministic network of comparable capacity trained with the same loss and protocol.","section":"Sec. 4.4, Table 3"},{"comment":"The diffusion loss is a direct L1 loss on the final reverse sample z_hat, rather than the standard variational noise-prediction objective used in DDPMs (Ho et al., Eq. 14). With T=8 and a 0.07M-parameter denoising network, it is not clear that the model learns the intermediate noise-conditioned distributions; it may instead learn an amortized mapping from Gaussian noise to the ARE residual. Please specify the noise schedule beta_t, give training or validation curves, and either justify Eq. (11) as a valid diffusion objective or describe the method as an iterative refinement rather than a diffusion model.","section":"Sec. 3.3, Eq. (11)"},{"comment":"The ARE defines the diffusion target z by a global mean/std alignment of amplitude spectra (Eq. 3). As the authors acknowledge in Supp. A7, this ignores local spatial variations in haze thickness. Because the diffusion target is derived entirely from this global alignment, any failure of the global-statistics assumption will propagate to the DM output regardless of the dehazing network's capacity. Please provide a quantitative evaluation on scenes with strongly non-uniform haze (for example, dense local fog regions) and discuss how the network's spatial processing compensates for the mis-specified target.","section":"Sec. 3.2, Eq. (3); Supp. A7"}],"minor_comments":[{"comment":"The abstract and contribution list claim 'the first work to apply the diffusion model to the unpaired image dehazing task,' but Sec. 2.2 cites DiffIR and HiDiff as DMs that reconstruct priors or latents for restoration. Please clarify the precise difference from these methods and restrict the novelty claim accordingly.","section":"Sec. 1, Contribution 1"},{"comment":"The description of the 'w/o DM' baseline is ambiguous: the sentence 'When we obtained z directly using the hazy image as the ARE's input' conflicts with Eq. (3), where ARE uses both hazy and clear amplitude spectra. Please specify the exact input configuration of the w/o DM baseline.","section":"Sec. 4.4, Table 3"},{"comment":"Table 2 reports FADE and BRISQUE without error bars; these no-reference metrics are known to be noisy, and the text describes FrDiff as 'favourable' on URHI while it is not the best. Please add standard deviations over multiple runs or bootstrap estimates and state explicitly whether lower values are better for both metrics.","section":"Sec. 4.1"},{"comment":"The L2 similarity between frequency features and GT is reported as ODCR:6.257 and FrDiff:2.156, but the feature definition, the normalization, and the frequency range used for the comparison are not specified; please provide the exact computation.","section":"Supp. A8, Fig. A4"},{"comment":"The training details specify an input patch size of 256x256 but do not state whether the SOTS-Indoor and SOTS-Outdoor evaluations are performed on full-resolution images or on cropped patches; please clarify the evaluation protocol to match the FLOPs computation in Table 1.","section":"Sec. 4.2"},{"comment":"The reference list contains malformed trailing page-number fields after some entries (for example, [1] ends with '5, 6, 3'), which appear to be citation locations leaked into the bibliography; please clean up the formatting.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The Table A2 versus Table 1 discrepancy is severe enough that the paper should not be accepted as is; I would advise the editor to request the corrected evaluation protocol and the release of code or weights before considering acceptance. The core idea is plausible, and the paper is generally well organized, but the empirical foundation of the headline claim is currently unresolved. The novelty claim of 'first DM for unpaired dehazing' should also be checked against prior unpaired dehazing diffusion works during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FrDiff is a genuinely new way to apply diffusion to unpaired dehazing: instead of generating full images, it reconstructs the amplitude residual between hazy and clear spectra, with a cheap phase-correction branch. That makes the paper worth reading, but the reported SOTS-Indoor numbers are internally inconsistent by about 9 dB (27.43 in Table 1 vs. 36.54 in Supp. Table A2), and the authors do not explain the gap.\n\nThe architecture is well thought out. The amplitude residual encoder is parameter-free, the phase correction module is inexpensive, and the ablation study shows each component has a role. The dark-channel statistics in the supplementary give a reasonable motivation for why amplitude exchange removes most haze. The main-table results beat the unpaired state of the art with fewer parameters, which is non-trivial.\n\nThe serious problem is that Table A2 reports 36.54 dB for the default λ=1 setting, while Table 1 says 27.43 dB. The sweep around that point is flat, so it is not a hyperparameter effect. Table 3's ablations use the 27.43 dB number, which makes the supplementary number look like a typo or a different protocol, but the paper does not say which. With no code or weights, the reader cannot verify either figure. This is not a minor typo; it affects the paper's central empirical claim.\n\nThe diffusion target being the ARE output is a design choice rather than a flaw – the DM learns a conditional mapper from hazy amplitude to residual – but the paper should be more explicit that it is a distillation of a fixed teacher. The global-statistics assumption is a real limitation, but it is already acknowledged in Supp A7.\n\nI would send this to peer review because the idea is novel and the presentation is mostly clear, but I would make the discrepancy a required revision. The paper is worth engaging with, but not in its current numerical state.","headline":"FrDiff is a novel take on unpaired dehazing, but its SOTS-Indoor numbers contradict each other by 9 dB and must be reconciled before the claims are trustworthy.","tokens_in":20614,"tokens_out":4349,"would_cite":false,"duration_ms":46110,"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":"FrDiff removes haze by having a diffusion model reconstruct the amplitude spectrum of clear images, and reports higher unpaired dehazing results on synthetic and real-world benchmarks than the compared baselines.","keywords":["unpaired image dehazing","diffusion model","frequency domain","amplitude spectrum","phase correction","image restoration","contrastive learning","amplitude residual encoder"],"falsifier":"Create synthetic hazy images with spatially varying transmission maps, such as thick haze near the top and thin haze near the bottom, and compare FrDiff's output to ground truth at a fixed average haze density; if its PSNR drops much more on these nonuniform images than on uniformly hazy images, the global amplitude-statistics assumption is the bottleneck.","tokens_in":19652,"feed_emoji":"🌫️","tokens_out":11070,"duration_ms":114232,"temperature":0.7,"pith_summary":"The paper presents FrDiff, described as the first diffusion model for unpaired image dehazing. Its claim is that haze degrades images mainly through the amplitude spectrum, the frequency component that carries illumination contrast, while the phase spectrum carries texture, so an unpaired dehazing model can remove haze by reconstructing only the amplitude spectrum of the clear domain. To do this it introduces an Amplitude Residual Encoder that aligns the global mean and standard deviation of a hazy image's amplitude to those of an unpaired clear image, and a Phase Correction Module that refines phase to suppress artifacts. On the RESIDE benchmarks it reports 27.43 dB PSNR and 0.957 SSIM on SOTS-Indoor and 22.75 dB and 0.914 SSIM on SOTS-Outdoor with 8.76M parameters and 54.56 GFLOPs. If the claim holds, diffusion models become a practical tool for unpaired restoration because the generative burden is moved from full images to a small frequency-domain residual.","feed_headline":"Denoise the amplitude spectrum, not the image, to dehaze","feed_subtitle":"FrDiff learns clear-image amplitude from unpaired data and beats contrastive baselines with 8.76M parameters.","key_machinery":"The load-bearing object is the amplitude residual $z = A'_h - A_h$, where $A_h$ is the amplitude spectrum of the hazy image and $A'_h = (\\sigma_c/\\sigma_h)(A_h - \\mu_h) + \\mu_c$ is the hazy amplitude after aligning its global mean and standard deviation to an unpaired clear amplitude spectrum. The Amplitude Residual Encoder computes $z$ in closed form with no learnable parameters; a conditional diffusion model with eight denoising steps generates $\\hat{z}$ from Gaussian noise conditioned on the hazy amplitude; frequency compensation layers inject $\\hat{z}$ into the dehazing network at multiple scales; and the Phase Correction Module uses a softmax-weighted convolution on the phase spectrum to remove artifacts. The residual formulation keeps the diffusion target small and illumination-only, which is why the method can work with few iterations.","core_discovery":"The central discovery is that unpaired image dehazing can be posed as a frequency-domain reconstruction task: instead of translating a hazy image into a clear image directly, the model generates the amplitude residual $z$ that closes the distribution gap between hazy and clear amplitudes. The paper's evidence is that swapping in a clear image's amplitude spectrum brings the dark-channel statistics of a hazy image close to those of true clear images, that the residual estimated by the Amplitude Residual Encoder provides supervision for diffusion training without extra parameters, and that the full FrDiff system outperforms the compared unpaired dehazing methods on SOTS-Indoor, SOTS-Outdoor, HSTS-Synth, and real-world benchmarks. The system does this with only eight denoising steps and a 0.07M-parameter denoiser attached to an 8.69M-parameter dehazing network.","pith_inferences":["An implication the paper leaves implicit is that the same two-stage diffusion-on-residual recipe could transfer to other unpaired restoration problems whose degradation is largely global, such as low-light enhancement, color-cast correction, or uniform blur.","Because the model without the diffusion component still aligns amplitudes but scores 24.66 dB, the closed-form alignment alone is not enough; the learned denoiser is what turns the residual into a working dehaze, suggesting a fair comparison against a non-diffusion regressor trained on the same residual.","A testable extension is spatially adaptive amplitude alignment; since the paper flags spatially varying haze as a limitation, showing that local or depth-conditional alignment improves real-world results would confirm that the global-statistics assumption, not network capacity, is the bottleneck.","The reported gain saturates as denoising steps grow from eight to thirty-two, hinting that a distilled single-step version could retain most of the benefit at a fraction of the FLOPs; the paper mentions one-step distillation as future acceleration."],"forward_implications":["Unpaired dehazing can be accomplished by a diffusion model producing only an amplitude residual, requiring about eight denoising steps rather than full image generation.","The reported SOTS-Indoor and SOTS-Outdoor results imply the approach beats cycle-consistency and contrastive-learning unpaired baselines at a lower parameter count of 8.76M.","The closed-form amplitude alignment supplies diffusion supervision without learnable parameters, so the training signal is tied directly to haze-related amplitude statistics.","The Phase Correction Module contributes a small but consistent gain, indicating that phase refinement is needed to remove artifacts left by amplitude-only reconstruction.","Ablation results attribute the largest single drop of 2.77 dB to removing the diffusion model, making the diffusion reconstruction the main driver of the reported performance."],"supporting_citations":[{"why":"Supplies the RESIDE training and test sets (ITS, SOTS-Indoor, SOTS-Outdoor, HSTS, URHI) on which all quantitative comparisons are run.","marker":"[28]"},{"why":"Supplies the dark channel prior used to justify the claim that haze is an amplitude-spectrum phenomenon.","marker":"[19]"},{"why":"Establishes the classical result that phase carries texture while amplitude carries illumination contrast, the premise for reconstructing amplitude.","marker":"[39]"},{"why":"Supports the same phase-versus-amplitude separation that motivates the Phase Correction Module.","marker":"[48]"},{"why":"Provides the denoising diffusion probabilistic model formalism on which the reconstruction of amplitude residuals is built.","marker":"[21]"},{"why":"Supplies the two-stage prior-representation diffusion training strategy that FrDiff adapts to the unpaired setting.","marker":"[59]"},{"why":"Supplies the NAFNet backbone used as the dehazing network and the frequency compensation injection points.","marker":"[6]"},{"why":"Supplies the patchwise contrastive loss that supervises the unpaired dehazing network together with the adversarial loss.","marker":"[40]"},{"why":"The strongest contrastive-learning unpaired dehazing baseline that FrDiff compares against and reports higher scores than.","marker":"[54]"}],"fun_headline_variants":["Diffuse the amplitude gap to dehaze unpaired images","FrDiff: amplitude-residual diffusion for unpaired dehazing","Unpaired dehazing via frequency-domain diffusion model","Generate clear amplitudes with diffusion to remove haze"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that haze is captured by a global shift in the amplitude spectrum, so aligning the overall mean and standard deviation of the hazy amplitude to an unpaired clear amplitude produces a residual that fully describes the haze; if real haze varies spatially across the image, this residual will be miscalibrated no matter how strong the diffusion model is.","fun_headline_variants_meta":{"raw":{"variants":["Diffuse the amplitude gap to dehaze unpaired images","FrDiff: amplitude-residual diffusion for unpaired dehazing","Unpaired dehazing via frequency-domain diffusion model","Generate clear amplitudes with diffusion to remove haze"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2704,"prompt_tokens":937,"completion_tokens":1767,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1704}},"tokens_in":553,"tokens_out":1767,"duration_ms":18128,"temperature":1.0,"reasoning_tokens":1704,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:56:22.261245+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Create synthetic hazy images with spatially varying transmission maps, such as thick haze near the top and thin haze near the bottom, and compare FrDiff's output to ground truth at a fixed average haze density; if its PSNR drops much more on these nonuniform images than on uniformly hazy images, the global amplitude-statistics assumption is the bottleneck.","supporting_citations":[{"cited_title":"Benchmarking single- image dehazing and beyond","cited_arxiv_id":null,"evidence_quote":"Supplies the RESIDE training and test sets (ITS, SOTS-Indoor, SOTS-Outdoor, HSTS, URHI) on which all quantitative comparisons are run."},{"cited_title":"Single image haze removal using dark channel prior","cited_arxiv_id":null,"evidence_quote":"Supplies the dark channel prior used to justify the claim that haze is an amplitude-spectrum phenomenon."},{"cited_title":"No-reference image quality assessment in the spa- tial domain","cited_arxiv_id":null,"evidence_quote":"Establishes the classical result that phase carries texture while amplitude carries illumination contrast, the premise for reconstructing amplitude."},{"cited_title":"Gated fusion net- work for single image dehazing","cited_arxiv_id":null,"evidence_quote":"Supports the same phase-versus-amplitude separation that motivates the Phase Correction Module."},{"cited_title":"Denoising diffu- sion probabilistic models","cited_arxiv_id":null,"evidence_quote":"Provides the denoising diffusion probabilistic model formalism on which the reconstruction of amplitude residuals is built."},{"cited_title":"DiffIR: Efficient diffusion model for image restoration","cited_arxiv_id":null,"evidence_quote":"Supplies the two-stage prior-representation diffusion training strategy that FrDiff adapts to the unpaired setting."},{"cited_title":"Contrastive learning for unpaired image-to-image translation","cited_arxiv_id":null,"evidence_quote":"Supplies the patchwise contrastive loss that supervises the unpaired dehazing network together with the adversarial loss."}],"review_version":1}