{"id":"30a7a653-719e-4d9e-a49e-a328077e4395","arxiv_id":"2501.12832","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A frequency-domain-guided diffusion framework improves restoration of hazy JPEG-compressed images, outperforming existing dehazing and cascade baselines on four datasets.","lead":"This paper presents an AI image-restoration method that cleans up hazy photos that have also been JPEG-compressed, a common real-world problem for traffic cameras. The method separates out compression damage and uses it to guide a diffusion model, reporting better quality than current dehazing tools on four benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I may compare FDG-Diff (trained on each compressed-haze split) against off-the-shelf baselines; a fair retrained-baseline test is needed before the SOTA claim is secure.","rationale":"The central claim is empirical: FDG-Diff consistently outperforms SOTA dehazing methods and FBCNN cascades on JPEG-compressed hazy images. The most load-bearing assumption for that claim is not the approximate spectral decomposition in Eq. 10/12; even if the multiplicative model is imperfect, the network is trained with ground-truth compressed/uncompressed pairs and can learn a nonlinear residual. The load-bearing condition is that the comparison is controlled. The paper never states that baselines are retrained on the same compressed-hazy training distribution, and the setup sentence in Section V-A only specifies the split for the proposed method. Off-the-shelf baselines would make the large Table I gaps unsurprising and not attributable to the proposed modules. This is testable by retraining the baselines under identical conditions. The reader's verdict is CONDITIONAL, and this concern also supports conditionality, so I do not move the verdict; I would keep it CONDITIONAL/UNCHANGED pending the retrained-baseline experiment. I disagree with the reader's choice of weakest_assumption, though I share the overall conditional assessment.","tokens_in":10118,"tokens_out":8163,"duration_ms":86262,"concrete_test":"Retrain every neural dehazing baseline and its FBCNN cascade on the exact QF80 compressed training split of I-Haze, O-Haze, Dense-Haze, and NH-Haze, using the same patch size, augmentation, optimizer, and iteration budget as FDG-Diff, then recompute all rows of Table I. If FDG-Diff still leads by comparable margins under this controlled protocol, the fairness concern is resolved; if the baselines close the gap, the headline outperformance claim must be weakened or re-scoped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-A states only that the last 10% of each dataset is reserved for testing and the rest for training, and then reports comparisons in Table I. It does not state that any neural dehazing baseline or FBCNN cascade is retrained or fine-tuned on the same QF80 compressed-hazy training split. If the baselines are used with published weights, the reported gains (e.g., NH-Haze PSNR 23.38 vs. 21.78 for FBCNN→DehazeDDPM) may reflect in-domain training of the proposed model rather than the HFCM/DADTP architecture. The reader's weakest assumption (the multiplicative model in Eq. 10/12) is less load-bearing because the spectrum decomposition network is trained directly with paired compressed/uncompressed hazy images, so the linear separation is only motivation, not an exact computational requirement. A controlled comparison with retrained baselines is therefore the pivotal check for the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FDG-Diff, a diffusion-based framework for dehazing JPEG-compressed hazy images. It consists of a DCT-domain spectrum decomposition network that separates a compression-effect spectrum from a corrected hazy image, a High-Frequency Compensation Module (HFCM) that uses wavelet-decomposed features and cross-attention with the compression spectrum, and a Degradation-Aware Denoising Timestep Predictor (DADTP) that adjusts per-patch diffusion timesteps based on a transmission map. Experiments on four hazy-image datasets compressed at QF 80 report consistent improvements over several dehazing baselines and their FBCNN cascades, with a component ablation on NH-Haze. The paper also derives a probabilistic argument that haze increases the chance of JPEG AC-coefficient annihilation.","tokens_in":10331,"tokens_out":9499,"duration_ms":97903,"significance":"The problem addressed is real and underexplored, and the proposed architecture is well motivated by the interaction between haze-induced low contrast and JPEG high-frequency loss. The derivation in Section III is correct under the stated uniform-transmission simplification and gives a concrete reason to expect joint degradation. The component ablation in Table II shows that both HFCM and DADTP contribute, and the paper includes a code link and clear figures. If the reported gains are confirmed under a controlled comparison with baselines trained on the same compressed data, this would be a valuable contribution to practical dehazing. The main risk is evaluation fairness, not the internal consistency of the method.","major_comments":[{"comment":"Section V-A and Table I: The paper does not state whether any of the neural dehazing baselines or FBCNN cascades were retrained or fine-tuned on the same QF-80 compressed-hazy training split that was used for FDG-Diff. If the baselines are evaluated with published weights, the reported improvements (e.g., NH-Haze PSNR 23.38 dB versus 21.78 dB for FBCNN→DehazeDDPM) could largely reflect in-domain training of the proposed model rather than the HFCM and DADTP components. This is the pivotal control for the central claim; please retrain all deep baselines and cascades under the same training data and protocol, or clearly report and justify any use of off-the-shelf weights.","section":"V-A, Table I"},{"comment":"Section V-A and Table I: Because the last 10% of each dataset is used for testing and the datasets contain only tens of images (I-Haze has about 30, O-Haze 45, Dense-Haze 33, NH-Haze 55), the test sets are likely to contain only 3–5 images. FID is known to be highly unstable at such sample sizes, yet FID is one of the four headline metrics in Table I and drives the claim of consistent perceptual improvement. The reported FID differences (e.g., 84.65 versus 121.06 on NH-Haze) may not be statistically meaningful. Please use a larger test split, report standard deviations or confidence intervals, or replace FID with a more stable distribution-level metric.","section":"V-A, Table I"}],"minor_comments":[{"comment":"The multiplicative matrix model for JPEG compression is not derived from the actual quantization and rounding process; the paper should state explicitly that this is a modeling approximation and that Eq. (12) holds only approximately.","section":"IV-A, Eqs. (10)–(12)"},{"comment":"It is unclear how patches with different predicted timesteps are trained and how the averaged noise estimate in Eq. (18) is used during sampling; please clarify the training loss for the DADTP and the fusion of overlapping patches with mismatched timesteps.","section":"IV-C, Eq. (18)"},{"comment":"The symbol 'hz(ν)' appears to be a typo for '\\hat f^z(ν)', and the sentence containing 'the the probability' should be corrected.","section":"III, Eq. (7)"},{"comment":"The experimental setup reports patch size, stride, and hardware, but not optimizer, learning rate, batch size, or number of training iterations for either network; please include these details or refer to the released code.","section":"V-A"},{"comment":"The ablation is performed only on NH-Haze; given that DADTP is motivated by spatially varying haze, reporting ablations on at least one non-uniform or dense-haze dataset would strengthen the claim.","section":"Table II"},{"comment":"The choice of QF=80 is justified by an informal web page; please justify it with a more standard source or by a small sensitivity study across quality factors.","section":"V-A, Reference [26]"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The main blocker is the fairness of the baseline comparison; I would ask the authors to retrain all competitors on the compressed-hazy training splits. The FID computation on tiny test sets is a secondary but real concern. I also noticed that reference [11] (Motion Mamba) is a self-citation used for a generic spectral-bias claim; the authors should cite a standard reference instead. The paper is within the journal's scope and the core idea is promising."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhi, if you have to desk-decide this one: it's worth sending out, but the central comparison needs a fix, not the architecture. The paper's real contribution is a new task setting, JPEG-compressed hazy image restoration, and a diffusion model built around it. That gap is genuine: compression is always present in real captures, and existing dehazing methods degrade on compressed inputs. I also think the analysis in Section III is a nice piece of motivation—the inequality that haze increases the probability of quantized AC coefficients going to zero is correct under their stated uniform-transmission simplification, and it does explain why dense haze regions lose high frequency first.\n\nWhat the paper does well: the spectrum decomposition network is trained directly against paired compressed/uncompressed hazy images, so Equation (10) is just a motivating linear model, not an exact computational requirement. The HFCM and DADTP are plausibly novel, and the ablation shows both contribute something. Results are consistent across four datasets on four metrics, and FID/LPIPS improvements over the higher-quality cascade baselines are sizable. The authors deserve credit for the transparent patch-based fusion formula in Eq. (18) that avoids edge artifacts.\n\nBut there are two soft spots. The first is the stress-test issue: Table I compares against baselines, presumably with published weights, while FDG-Diff is trained on the QF-80 compressed training split. No retraining or fine-tuning of baselines is described. That alone could explain a large share of the reported gains. I'd want to see the best cascade baselines (especially FBCNN→DehazeDDPM) retrained or at least fine-tuned on the same compressed training data before believing the SOTA claim. The second is the experimental reporting: no error bars, no multiple seeds, one QF (80), and one train/test split. That's minor in this subfield, but it does cap the confidence of the headline numbers.\n\nThe 'first learning-based' claim is probably unverifiable—there is prior learning-based JPEG artifact removal and dehazing, but joint compressed-haze learning may well be new; I'd soften it to 'to our knowledge.' The citation to their own Motion Mamba paper for spectral bias is odd, but not a flaw—Rahaman et al. is also cited.\n\nBottom line: the method is coherent, the gap is real, and the modules earn their keep in the ablation. This deserves a serious referee, with a request for retrained baselines and error bars. I wouldn't block on the approximation in Eq. (10).","headline":"A solid, well-engineered entry into a real practical gap—compressed hazy image restoration—but the SOTA claim rests on a comparison against off-the-shelf baselines, so it needs a retrained-baseline check before it is secure.","tokens_in":10832,"tokens_out":2131,"would_cite":false,"duration_ms":21760,"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":"This paper proposes FDG-Diff, a diffusion-based framework that restores JPEG-compressed hazy images by first separating compression loss in the DCT domain, then using that frequency information to guide high-frequency compensation and…","keywords":["image dehazing","JPEG compression","diffusion model","frequency-domain guidance","DCT spectrum decomposition","high-frequency compensation","degradation-aware timestep prediction","compressed hazy image restoration"],"falsifier":"Train the spectrum decomposition network on hazy images compressed at QF 10 and compare the predicted compression spectrum against the true rounding residual; if the predicted spectrum correlates with scene content (e.g., contains edges of the hazy scene) rather than only quantizer structure, then the content-independent $\\varphi$ assumption fails and the cross-attention guidance is partly using leaked image content. A second check: if removing HFCM or DADTP and simply training the DDPM on corrected hazy images at QF 10 closes the gap to FDG-Diff, then the claimed mechanism is not the driver.","tokens_in":9900,"feed_emoji":"🌋","tokens_out":4593,"duration_ms":43318,"temperature":0.7,"pith_summary":"This paper argues that JPEG compression and haze act jointly, with haze making high-frequency DCT coefficients more likely to be zeroed during quantization. On that basis it proposes FDG-Diff, a diffusion-based dehazing framework that first separates the compression effect from the hazy image in the DCT domain, then uses the predicted compression spectrum to guide high-frequency compensation during reverse sampling. A separate module predicts a per-patch denoising timestep from the transmission map, letting dense-haze regions receive stronger restoration. Across four compressed dehazing benchmarks at quality factor 80, the method reports consistently higher PSNR and SSIM and lower LPIPS and FID than state-of-the-art dehazing models and their cascades with the FBCNN JPEG artifact remover.","feed_headline":"Frequency-guided diffusion beats cascades on JPEG-degraded hazy images","feed_subtitle":"Separating compression loss in the DCT domain lifts NH-Haze PSNR to 23.38 dB at QF 80, beyond the best artifact-removal cascade.","key_machinery":"The carrying mechanism is the spectrum decomposition network, which models JPEG compression as a content-independent multiplicative matrix $\\varphi$ in the DCT domain ($I_c = \\varphi^T I$), then uses log and DCT to turn the product into a sum of a compression-effect spectrum and a lossless hazy-image spectrum. This predicted compression spectrum is the key that unlocks the other two components: the High-Frequency Compensation Module uses Haar-wavelet sampling of skip-connection features and cross-attention against the compression spectrum to add back the high frequencies the quantizer removed, while the Degradation-Aware Denoising Timestep Predictor maps a dark-channel transmission estimate to a per-patch timestep offset, so denser haze gets stronger denoising. The DDPM with these components is trained with an L1 noise-prediction loss and uses overlapping sliding patches with averaged noise estimates to avoid seam artifacts.","core_discovery":"The central discovery is that the joint loss from haze and JPEG compression is not merely additive: because haze attenuates AC coefficients before quantization, compressed hazy images lose high-frequency structure far more aggressively than either degradation alone, and this loss is concentrated in dense-haze regions. FDG-Diff exploits this by decomposing the compressed hazy image in the DCT domain into a compression-effect spectrum and a corrected hazy image, with the multiplicative compression model turned additive via a log transform. The corrected image conditions a patch-based DDPM, while the compression spectrum is fed through cross-attention in the High-Frequency Compensation Module to restore wavelet-extracted high-frequency features. The Degradation-Aware Denoising Timestep Predictor adds region-specific denoising intensity. The reported result is consistent state-of-the-art improvement, e.g. NH-Haze at QF 80: PSNR 23.38 dB, SSIM 0.81, LPIPS 0.14, FID 84.65, against 21.78 dB, 0.72, 0.17, 121.06 for the best cascade (FBCNN followed by DehazeDDPM).","pith_inferences":["Because the paper shows the joint loss grows with haze density, one testable extension is that FDG-Diff's margin over cascades should widen on datasets with denser or more non-uniform haze; the reported Dense-Haze and NH-Haze gains are consistent with that.","The content-independence of $\\varphi$ is an idealization; at low quality factors where JPEG rounding and zeroing become highly nonlinear, the predicted compression spectrum may carry residual scene content, so the method's advantage may shrink and a content-dependent correction would be needed.","DADTP is not limited to transmission maps; any regionwise degradation estimate (e.g., local noise variance or blockiness map) could drive timestep offsets in other restoration tasks, suggesting a general region-adaptive diffusion scheduling idea.","The single-channel Y analysis in the problem statement leaves chrominance largely unmodelled; extending the decomposition to the full YUV structure could further improve color fidelity, which the visual results already hint at."],"forward_implications":["Existing dehazing models that ignore compression lose several dB on JPEG inputs, so compression-awareness should be a default consideration for real-world dehazing deployments.","Cascading a JPEG artifact remover before dehazing outperforms the reverse order, but the paper's best cascade still trails FDG-Diff, suggesting that explicit joint modelling beats serial de-artifacting.","The frequency-domain separation idea transfers: any degradation that is approximately multiplicative in a transform domain could be decomposed and used as diffusion guidance.","DADTP's transmission-based timestep prediction gives a general recipe for spatially varying degradations beyond haze, such as non-uniform blur or local compression artifacts."],"supporting_citations":[{"why":"Supplies the dark channel prior transmission map used by DADTP to quantify regional degradation, and serves as a baseline dehazing method.","marker":"[1]"},{"why":"Provides the denoising diffusion probabilistic model formulation that FDG-Diff extends with frequency guidance and adaptive timesteps.","marker":"[10]"},{"why":"Documents the spectral bias of neural networks, motivating the need for explicit high-frequency compensation in the HFCM.","marker":"[12]"},{"why":"FBCNN is the JPEG artifact removal model used to build the cascaded baselines that FDG-Diff must outperform.","marker":"[18]"},{"why":"Demonstrates modelling JPEG compression effects directly in the DCT domain, underpinning the spectrum decomposition network design.","marker":"[20]"},{"why":"I-Haze is one of the four real hazy benchmarks used to evaluate FDG-Diff against state-of-the-art methods.","marker":"[8]"},{"why":"O-Haze provides the outdoor hazy benchmark used in the quantitative comparison.","marker":"[22]"},{"why":"Dense-Haze provides the dense-haze benchmark where joint loss is most severe and FDG-Diff reports its largest gains.","marker":"[23]"},{"why":"NH-Haze provides the non-homogeneous haze benchmark used for the headline results and the ablation study.","marker":"[24]"}],"fun_headline_variants":["Frequency-guided diffusion beats cascades on compressed hazy images","DCT-domain diffusion restores JPEG-hazy images beyond cascades","Frequency-guidance diffusion sets SOTA on compressed hazy images","Diffusion with DCT guidance beats artifact-removal cascades"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that JPEG compression acts as a content-independent multiplicative matrix, so after a log transform and DCT the compression effect and the clean hazy image simply add; real JPEG quantization rounds and zeroes coefficients depending on the image content and haze density, so the separation is approximate and the predicted compression spectrum may carry residual scene information.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-guided diffusion beats cascades on compressed hazy images","DCT-domain diffusion restores JPEG-hazy images beyond cascades","Frequency-guidance diffusion sets SOTA on compressed hazy images","Diffusion with DCT guidance beats artifact-removal cascades"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001081,"raw_usage":{"total_tokens":4539,"prompt_tokens":981,"completion_tokens":3558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":3485}},"tokens_in":597,"tokens_out":3558,"duration_ms":27269,"temperature":1.0,"reasoning_tokens":3485,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:43:30.377113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the spectrum decomposition network on hazy images compressed at QF 10 and compare the predicted compression spectrum against the true rounding residual; if the predicted spectrum correlates with scene content (e.g., contains edges of the hazy scene) rather than only quantizer structure, then the content-independent $\\varphi$ assumption fails and the cross-attention guidance is partly using leaked image content. A second check: if removing HFCM or DADTP and simply training the DDPM on corrected hazy images at QF 10 closes the gap to FDG-Diff, then the claimed mechanism is not the driver.","supporting_citations":[{"cited_title":"Single image haze removal using dark channel prior,","cited_arxiv_id":null,"evidence_quote":"Supplies the dark channel prior transmission map used by DADTP to quantify regional degradation, and serves as a baseline dehazing method."},{"cited_title":"Denoising diffusion probabilistic models.,","cited_arxiv_id":null,"evidence_quote":"Provides the denoising diffusion probabilistic model formulation that FDG-Diff extends with frequency guidance and adaptive timesteps."},{"cited_title":"On the spectral bias of neural networks,","cited_arxiv_id":null,"evidence_quote":"Documents the spectral bias of neural networks, motivating the need for explicit high-frequency compensation in the HFCM."},{"cited_title":"Towards flexible blind jpeg artifacts removal,","cited_arxiv_id":null,"evidence_quote":"FBCNN is the JPEG artifact removal model used to build the cascaded baselines that FDG-Diff must outperform."},{"cited_title":"Jpeg quantized coefficient recovery via dct domain spatial-frequential transformer,","cited_arxiv_id":null,"evidence_quote":"Demonstrates modelling JPEG compression effects directly in the DCT domain, underpinning the spectrum decomposition network design."},{"cited_title":"I-haze: A dehazing benchmark with real hazy and haze-free indoor images,","cited_arxiv_id":null,"evidence_quote":"I-Haze is one of the four real hazy benchmarks used to evaluate FDG-Diff against state-of-the-art methods."},{"cited_title":"O-haze: A dehazing benchmark with real hazy and haze-free outdoor images,","cited_arxiv_id":null,"evidence_quote":"O-Haze provides the outdoor hazy benchmark used in the quantitative comparison."},{"cited_title":"Dense-haze: A benchmark for image dehazing with dense-haze and haze-free images,","cited_arxiv_id":null,"evidence_quote":"Dense-Haze provides the dense-haze benchmark where joint loss is most severe and FDG-Diff reports its largest gains."},{"cited_title":"Nh-haze: An image dehazing benchmark with non-homogeneous hazy and haze-free images,","cited_arxiv_id":null,"evidence_quote":"NH-Haze provides the non-homogeneous haze benchmark used for the headline results and the ablation study."}],"review_version":1}