{"id":"2c5abc25-0e59-48b7-8c31-78f168a98a4b","arxiv_id":"2411.13961","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A zero-shot low-light image enhancement method that injects joint wavelet and Fourier frequency priors into a pre-trained ImageNet diffusion model, reporting top zero-shot metrics on LOL and SICE.","lead":"This paper presents a zero-shot method for brightening low-light photos, using a pre-trained diffusion model guided by wavelet and Fourier frequency information from the input image. It reports better enhancement quality than several existing zero-shot methods on the LOL and SICE benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The structural prior is taken from the input at every sampling step, so the method cannot recover corrupted high-frequency details and its severe-degradation claim is unverified.","rationale":"The reader's weakest_assumption identifies exactly this issue: the input-derived Fourier phase and wavelet high-frequency coefficients are treated as reliable structural priors at every sampling step, so input corruption is inherited. This is the most load-bearing concern because it attacks the method's core novelty (joint frequency-domain priors) and its stated motivation of handling severe unknown degradation. It is not refuted by the benchmark numbers, which only show that on relatively clean low-light datasets the approach is effective. The missing code and missing comparison to prior work are secondary: they hinder reproducibility but do not as directly expose a mechanism-level limitation. Our proposed synthetic-noise test would settle whether the structural prior actually recovers details or merely copies them from the input. Since the concern matches the reader's and the reader already issued a conditional verdict, no verdict change is needed.","tokens_in":7872,"tokens_out":9445,"duration_ms":94758,"concrete_test":"Construct a synthetic severe-degradation test: take clean images from the LOL test set, add Gaussian noise (e.g., σ=25) or a small blur kernel to the low-light inputs, then run the proposed method and FourierDiff on the same degraded inputs. Compute the wavelet high-frequency (HH) subband energy of the output and compare it with the input's HH energy and the clean ground truth's HH energy. If the output's HH energy closely tracks the input's HH energy (e.g., correlation >0.9) rather than the clean's, then Eq. (9) is copying the degraded high-frequency content and the method does not repair corrupted structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism (Eqs. 4-9) fixes the input's wavelet high-frequency coefficients H2_L and Fourier phase phaL into every inverse diffusion step. The diffusion process operates only on the first-level low-frequency subband LL (Eq. 4), so it never regenerates the structural detail bands; those bands are overwritten with the input's values. If the low-light input's phase or high-frequency content is corrupted by noise, blur, or compression—precisely the 'unknown severe degradation' the abstract says the method addresses—those corruptions are injected into the output at every step and cannot be repaired by the diffusion prior. The final denoise module (Eq. 11) is a post-hoc, unspecified simplification, and Table I contains no experiment on noisy or blurred inputs. The strong LOL/SICE scores may therefore reflect the relative cleanliness of those datasets' high-frequency content rather than an ability to recover lost structure, leaving the paper's central robustness claim unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a zero-shot low-light image enhancement method that combines wavelet and Fourier frequency-domain priors with a pre-trained unconditional diffusion model. The method performs the diffusion sampling in the wavelet low-frequency subband of the input, and at each reverse step replaces the sample's wavelet high-frequency coefficients and Fourier phase with those of the input while updating the Fourier amplitude with a learned luminance factor (Eqs. (4)-(9)). It adds CLIP-based text guidance and a non-reference brightness loss to optimise the luminance factor, and ends with an unspecified denoising post-processing step (Eq. (11)). The authors report state-of-the-art results among zero-shot methods on the LOL and SICE benchmarks, with an ablation study on LOL.","tokens_in":8161,"tokens_out":3822,"duration_ms":40365,"significance":"If the reported results are reproducible, this is a potentially useful contribution: it achieves zero-shot low-light enhancement without paired training data, and it gives a concrete way to combine wavelet and Fourier priors with diffusion sampling. The paper ships a clear pipeline, compares against many unsupervised methods, and promises public code, all of which are strengths. However, the significance is conditional: the central mechanism fixes the input's high-frequency and phase information at every sampling step, which means the method cannot repair corrupted or lost structural details; the claimed robustness to 'unknown severe degradation' is not tested on noisy, blurred, or compressed inputs. The validation is also thin: Table I reports only single-run metrics with no variance or statistical testing, and the closest published method from the same group, reference [31], is not compared. I do not see a circularity problem in the use of non-reference losses, since L_bri and L_TG are distinct from the ground-truth PSNR/SSIM/LPIPS/FID evaluation criteria.","major_comments":[{"comment":"The structural-prior injection cannot recover corrupted high-frequency detail. The diffusion process operates only on the wavelet low-frequency subband LL (Eq. (4)), and in Eq. (9) the sample's high-frequency coefficients H2_L and Fourier phase phaL are taken directly from the input image. If the input's phase or high-frequency content is corrupted by noise, blur, or compression, those corruptions are copied into the output at every reverse step and cannot be repaired by the diffusion prior. The abstract's claim about handling 'unknown severe degradation' is therefore unsupported, and Table I contains no experiment with noisy, blurred, or compressed inputs. Please either add such experiments or substantially temper the robustness claim.","section":"Sec. II-B, Eq. (9)"},{"comment":"The use of the DDNM-style joint-distribution update is not justified. The formula in Eq. (10) is derived for the linear measurement model y = Ax, where the corrected sample lies in the range of the measurement operator. Here x1_t is produced by a nonlinear sequence of DWT, FFT, amplitude/phase replacement, IFFT, and IDWT operations, so the measurement is not linear and the assumptions behind Eq. (10) do not automatically hold. The authors should either prove that the update remains valid for this nonlinear projection or provide an empirical validation, since this equation is load-bearing for the entire sampling algorithm.","section":"Sec. II-B, Eq. (10)"},{"comment":"The quantitative validation is too thin. All metrics in Table I are reported as single-run point estimates with no standard deviation, number of runs, or statistical significance tests, which matters because diffusion sampling is stochastic and the method includes per-image optimisation. In addition, the closest prior work from the same group, 'Low-Light Image Enhancement via CLIP-Fourier Guided Wavelet Diffusion' (reference [31]), is not included as a baseline; given its apparent similarity to the proposed pipeline, omitting it weakens the claim of state-of-the-art performance among zero-shot methods. Please add repeated-run statistics and a direct comparison to [31].","section":"Sec. III-A, Table I"},{"comment":"Key hyperparameters and implementation details are missing, which prevents reproduction. The brightness level E in Eq. (13) is never given; the alternating optimisation interval S is set to 200 but its exact use during sampling is not described; the optimiser and learning rate for ϑ are unspecified; the exact CLIP prompts Tp and Tn are not stated; and the 'simplified' denoising module in Eq. (11) is not described beyond a reference to [17], [24]. Please specify these details or provide the code in a form that allows the experiments to be reproduced.","section":"Sec. II-C and Sec. III-A.2"}],"minor_comments":[{"comment":"The notation around Eq. (1) is garbled: 'αt=Qt i=1 αi' should be the product notation, and 't ∈ {[1,...T}' should be 't ∈ {1,...,T}'.","section":"Sec. II-A"},{"comment":"The section heading contains a typo: 'F ourier' should be 'Fourier'.","section":"Sec. II-B heading"},{"comment":"The text guidance loss sums over t ∈ [0,T], but with T=1000 this would require CLIP evaluations at every step. Please clarify how the loss is actually applied in the alternating-optimisation schedule with interval S.","section":"Sec. II-C, Eq. (12)"},{"comment":"The notation L2_L and H2_L is confusing because the superscript 2 could be read as a power rather than as a label for the second-level wavelet decomposition. Please define the notation explicitly and consider using different symbols.","section":"Sec. II-B, Eqs. (4)-(9)"},{"comment":"The visual comparison figure is difficult to read because the method labels are placed directly on the images and some are partially obscured; please provide a clearer layout with separate labels.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The omission of the authors' own closely related reference [31] from the comparisons is surprising and should be addressed in revision; it is relevant both to the novelty claim and to the fairness of the state-of-the-art comparison. The central mechanism may be sound on clean low-light inputs, but the robustness claim and the theoretical justification of Eq. (10) need more work before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible zero-shot low-light enhancement method with a clean pipeline and strong reported numbers, but it is a close sibling of the authors' own prior work [31] and never compares to it. The abstract's claim about handling \"unknown severe degradation\" is not backed by the experiments.\n\nWhat is new: the diffusion process runs on the first-level wavelet low-frequency subband, and at each sampling step the method substitutes in Fourier amplitude from the input's second-level low-frequency subband, plus the input's Fourier phase and high-frequency wavelet coefficients. A learnable brightness factor and a CLIP text-similarity loss steer the sampling. This is a sensible compositional design, and the ablations show the wavelet prior helps and text guidance adds a bit. On LOL, the reported zero-shot numbers are the best among the listed methods (PSNR 20.9 vs FourierDiff's 18.7, FID 63.6 vs 86.5). The pipeline is easy to follow and the writing is mostly clear.\n\nSoft spots, in rough order of importance. First, no comparison to the authors' own [31], which appears to be the direct predecessor (CLIP-Fourier guided wavelet diffusion). Without that comparison, the marginal contribution of this paper is unclear. Second, the paper reports single-run metrics with no error bars or statistical tests; diffusion sampling is stochastic, so this matters. Third, several implementation details are missing: the denoising module in Eq. 11 is described only as \"simplified,\" the optimizer for the learnable luminance factor is not stated, the exact CLIP prompts are not given, and the brightness level E is unspecified. Fourth, and structurally important: the method permanently injects the input's phase and high-frequency wavelet coefficients into every sampling step. If those coefficients are corrupted by noise, blur, or compression, the diffusion prior cannot repair them. The paper claims robustness to severe unknown degradation but offers no experiments on noisy or blurred inputs; LOL and SICE are relatively clean in their high-frequency content. That claim is unsupported. Fifth, the code is promised but not provided, so the numbers are not independently checkable.\n\nWho this is for: researchers working on zero-shot diffusion-based restoration. They will find the frequency-domain composition worth reading as a baseline. The paper deserves a serious referee, but it needs a revision that adds the [31] comparison, error bars, implementation details, and at least one experiment with corrupted high-frequency content before the strong claims can be trusted.","headline":"A plausible zero-shot low-light enhancement method with strong reported numbers, but it is a close sibling of the authors' own [31] and the severe-degradation claim is untested.","tokens_in":8621,"tokens_out":2427,"would_cite":false,"duration_ms":26879,"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":"A joint wavelet-Fourier prior guides zero-shot diffusion low-light enhancement.","keywords":["zero-shot low-light enhancement","diffusion model","wavelet transform","Fourier transform","frequency-domain prior","CLIP text guidance","illumination prior"],"falsifier":"Take a low-light image, add realistic sensor noise or moderate JPEG compression to the dark input, run the method, and compare the output with a clean ground truth; if the PSNR and SSIM gains over simpler baselines shrink or reverse while the method wins on clean synthetic low-light images, the prior-injection point of failure is confirmed.","tokens_in":7703,"feed_emoji":"💡","tokens_out":7902,"duration_ms":73984,"temperature":0.7,"pith_summary":"The paper attempts to show that the weak point of zero-shot low-light enhancement with a pretrained diffusion model is the absence of image-specific illumination and structure information, and that a rich prior obtained by jointly decomposing the input in the wavelet and Fourier frequency domains can supply that information at every reverse-sampling step without any paired training data. It claims that running the diffusion process in the wavelet low-frequency band and continuously fusing the sample with the input's Fourier phase and high-frequency wavelet coefficients keeps the output faithful to the original scene while correcting its exposure. The reported numbers place the method ahead of the compared zero-shot diffusion baselines on LOL, SICE, and unpaired scenes, with best zero-shot LPIPS 0.281 and FID 63.601 on LOL. If the claim holds, zero-shot enhancement becomes a competitive alternative to unpaired-training approaches in complex real-world lighting.","feed_headline":"Frequency-domain priors lift zero-shot low-light enhancement","feed_subtitle":"Grafting the input's phase and high-frequency detail into each sampling step lifts perception scores without paired data.","key_machinery":"The load-bearing object is the joint wavelet-Fourier frequency-domain prior. A two-level discrete wavelet transform of the low-light input yields a low-frequency band ($L_L$, then $L^2_L$) that concentrates illumination, and high-frequency bands ($H_L$, $H^2_L$) that concentrate structure; the Fourier transform of the low-frequency bands separates amplitude (illumination) from phase (structure). At each denoising step the algorithm combines the sampling result's amplitude with the input-derived amplitude, replaces the phase and high-frequency coefficients with the input's, and inverts the transforms, so the updated sample is steered both toward correct exposure and toward the input's content. The second mechanism is the null-space sampling rule from Eq. (10), which keeps the trajectory consistent with the input while accepting the frequency-domain edits, plus a CLIP text loss and a brightness loss that optimise the learned blend factor.","core_discovery":"The paper's central claim is that zero-shot low-light enhancement with a pretrained diffusion model fails not because the diffusion prior is weak but because it lacks image-specific illumination and structure guidance, and that a prior assembled from the joint wavelet and Fourier decomposition of the input supplies exactly that missing guidance. The method runs the denoising process on the wavelet low-frequency subband of the input and, at every inverse step, rebuilds the sample from the input's Fourier phase and second-level wavelet high-frequency coefficients, with an amplitude formed by a learned blend of the sample's amplitude and the input's amplitude (Eq. (9)). The rebuilt sample is then merged with the denoiser's prediction using the null-space sampling rule (Eq. (10)), and the final output is inverse-wavelet-transformed with the input's high-frequency band and refined by a simple denoiser. On the LOL and SICE benchmarks plus an unpaired set, the paper reports the best perception scores among the compared zero-shot methods, with the best overall LPIPS and FID on LOL.","pith_inferences":["Editorial inference: because the prior freezes the degraded input's phase and high-frequency coefficients, the method should inherit whatever defects those coefficients contain; adding realistic sensor noise or JPEG blocking to a dark input should measurably erode the LPIPS/FID gains, a test the paper does not run.","Editorial inference: the same grafting scheme is not tied to light enhancement; swapping the target coefficients should port it to dehazing, deblurring, or super-resolution, where the input's low-frequency structure is also more reliable than its corrupted high frequencies.","Editorial inference: the learnable luminance factor is a single global scalar in the paper, so applying the method to scenes with strongly non-uniform illumination would likely benefit from a spatially varying version of $\\vartheta$; the paper's fixed brightness level in Eq. (13) does not model local lighting."],"forward_implications":["Zero-shot can rival unpaired-training enhancers on standard benchmarks: on LOL the method reports PSNR 20.922, SSIM 0.811, LPIPS 0.281, and FID 63.601, the top zero-shot row in Table I.","Because the diffusion prior is frozen and only frequency coefficients and a scalar factor are changed, the method generalises to new scenes without retraining, as shown by results on SICE and on LIME, DICM, and MEF unpaired images.","Pinning the sample's phase and high-frequency wavelet band to the input keeps the output structurally consistent with the source image, which addresses the color distortion and random-detail artifacts the paper attributes to GDP and FourierDiff.","The ablation in Table II indicates the wavelet branch carries most of the gain; the CLIP text term mainly refines SSIM and perceptual scores rather than brightness."],"supporting_citations":[{"why":"Shows that the wavelet low-frequency band carries a better exposure effect than the degraded image, motivating the transfer of the diffusion process into the low-frequency domain.","marker":"[10]"},{"why":"The Fourier-only zero-shot diffusion baseline this work extends; it supplies the sampling-guidance setup into which the wavelet branch is added.","marker":"[20]"},{"why":"Demonstrates that high-frequency information is easily damaged during diffusion, supporting the decision to replace the sample's high-frequency coefficients and phase with the input's.","marker":"[26]"},{"why":"Provides the null-space sampling rule (Eq. (10)) used to merge the frequency-updated sample with the denoiser output at each step.","marker":"[28]"},{"why":"Supplies the frozen CLIP text and image encoders used for positive/negative prompt guidance in the sampling loss.","marker":"[25]"},{"why":"Provide the denoising diffusion post-processing and intensity-aware lookup-table ideas simplified into the final refinement step (Eq. (11)).","marker":"[17], [24]"},{"why":"Provides the LOL paired test set used for the quantitative comparison in Table I.","marker":"[30]"}],"fun_headline_variants":["Zero-shot low-light fix via wavelet-Fourier diffusion priors","Joint frequency priors brighten zero-shot low-light diffusion","Wavelet+Fourier priors lift low-light zero-shot perception","No paired data? Frequency priors guide zero-shot low-light","Frequency priors power zero-shot low-light enhancement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the fine details and phase information inherited from the dark input photo are accurate and clean, not ruined by noise or compression; if they are ruined, the method keeps reinserting those flaws at every sampling step and cannot repair them.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot low-light fix via wavelet-Fourier diffusion priors","Joint frequency priors brighten zero-shot low-light diffusion","Wavelet+Fourier priors lift low-light zero-shot perception","No paired data? Frequency priors guide zero-shot low-light","Frequency priors power zero-shot low-light enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000864,"raw_usage":{"total_tokens":3760,"prompt_tokens":975,"completion_tokens":2785,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":2701}},"tokens_in":591,"tokens_out":2785,"duration_ms":19382,"temperature":1.0,"reasoning_tokens":2701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:41:42.062845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a low-light image, add realistic sensor noise or moderate JPEG compression to the dark input, run the method, and compare the output with a clean ground truth; if the PSNR and SSIM gains over simpler baselines shrink or reverse while the method wins on clean synthetic low-light images, the prior-injection point of failure is confirmed.","supporting_citations":[{"cited_title":"Fourier priors-guided diffusion for zero-shot joint low-light enhancement and deblurring","cited_arxiv_id":null,"evidence_quote":"The Fourier-only zero-shot diffusion baseline this work extends; it supplies the sampling-guidance setup into which the wavelet branch is added."},{"cited_title":"Freeu: Free lunch in diffusion u-net","cited_arxiv_id":null,"evidence_quote":"Demonstrates that high-frequency information is easily damaged during diffusion, supporting the decision to replace the sample's high-frequency coefficients and phase with the input's."}],"review_version":1}