{"id":"0d74172a-6cbd-4684-9b69-d6a2537b77d1","arxiv_id":"2501.12860","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CrossDiff applies a diffusion probabilistic model with a cross-conditional encoder-decoder to crack segmentation, reporting state-of-the-art Dice and IoU scores on five datasets.","lead":"This paper introduces CrossDiff, a diffusion-based model with a cross-conditional encoder-decoder, for segmenting cracks in concrete surface images. The authors report gains of about 8 percentage points in Dice score and IoU over prior state-of-the-art methods across five crack datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8% SOTA claim hinges on an unspecified train/test split of the combined 7169-image dataset; a leak could explain most of the gain.","rationale":"I read the paper as a plausible architectural contribution that is being evaluated on the strength of an empirical SOTA claim. The most load-bearing condition for that claim is a fair, leakage-free comparison, and the paper does not document the train/test split of the combined dataset that contains the five evaluation sets. My emphasis differs from the reader's: the text explicitly says 'All models are set 5 times of ensemble in the inference', which weakens the reader's primary ensemble-asymmetry concern, whereas the undefined split is unambiguous and potentially more damaging. If the split is a random partition of the combined Li et al. set, images from the same original dataset can straddle train and test, and because crack datasets contain near-duplicate patches from the same surfaces, this can inflate scores substantially. The ablation table limits the attribution of the headline gain to the novel cross modules: they add only 1.36% and 1.35% average IoU, so a large unexplained gap remains. I am not claiming the method is fraudulent or that the result is certainly wrong; the concern is that the evidence as written cannot distinguish architectural superiority from leakage or protocol differences. The paper has some independent support in the form of a structurally described architecture and an ablation table, but code is not released and no error bars are given. Conditional acceptance is therefore the appropriate verdict: the central claim should not be rejected out of hand, but it cannot be accepted as stated until the split and inference protocol are clarified and the comparison is rerun under matching conditions. If the disclosed split and rerun show a clean 8-point gain, the concern is resolved; if not, the verdict should move toward rejection or unverified.","tokens_in":10679,"tokens_out":5138,"duration_ms":56645,"concrete_test":"Request the exact partition used for Table 1: number of training/validation/test images per dataset, with image IDs, and confirm that no test image appears in training. Then independently retrain CrossDiff and SegDecNet++ on the same split with identical inference protocols (single model and 5-run STAPLE ensemble) and recompute the average Dice and IoU. If the average gain over SegDecNet++ is no longer about 8 points but falls toward the roughly 2.7-point IoU ablation contribution of the cross modules, the headline result is an artifact of leakage or protocol differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Dataset section states that the model is trained and evaluated on the 7169-image compilation by Li et al. (2021) assembled from CFD, CrackTree200, DeepCrack, GAPs384, Rissbilder, CRACK500, and Non-crack, but it never describes how these 7169 images are partitioned into train/test sets. Table 1 then reports per-dataset Dice/IoU for the five constituent datasets. If, as written, the same combined set is used for training and for those evaluations, or if the partition is a random image-level split of the combined set, images from the same source dataset (and often the same pavement surface) can appear in both training and test, producing near-duplicate leakage. This is not a hypothetical: crack datasets are captured as image patches from a small number of scenes, so image-level leakage is easy to trigger. The magnitude matters because the cross-encoder and cross-decoder ablate to only +2.71 average IoU (Table 3: 68.90 to 71.61), so the remaining roughly six percentage points of the 8-point average gain over SegDecNet++ must come from either the diffusion backbone, ensembling/STAPLE, or the evaluation setup. The architecture may be sound, but the headline claim is not independently checkable until the split is specified and the test images are shown to be disjoint from training. The sentence in Implementation Details that 'all models are set 5 times of ensemble in the inference' also needs elaboration, because STAPLE is a label-fusion algorithm and it is unclear how it is applied to discriminative baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes CrossDiff, a diffusion probabilistic model with a cross-conditional encoder-decoder for pixel-level crack segmentation. The method combines a DDPM backbone with a ViT-based Cross Encoder, a Fusion Nexus module, and a training-only Cross Decoder, and is trained on a 7169-image compilation of five crack datasets. The paper reports Dice and IoU for five datasets, compares against DeepLabv3+, SCCDNet-D32, DeepCrack, CrackFormer, and SegDecNet++, and claims an average 8.0% improvement over SegDecNet++. An ablation study examines the contribution of the Cross Encoder and Cross Decoder. The central claim is empirical state-of-the-art performance for slender crack segmentation.","tokens_in":10942,"tokens_out":7736,"duration_ms":75556,"significance":"If the reported results are reproducible under a clean evaluation protocol, CrossDiff would be a useful new application of diffusion probabilistic models to slender-structure segmentation and a credible baseline for crack segmentation. The paper has clear strengths: it addresses a practically important problem, uses standard metrics (Dice and IoU), compares with several recent named baselines, and includes an ablation. However, the current evidence does not establish the central claim because the training/test split and the inference protocol are underspecified, and the ablation table is difficult to interpret. The contribution is therefore promising but not yet verified.","major_comments":[{"comment":"The Dataset section states that the model is trained and evaluated on the 7169-image compilation of Li et al. (2021), but it does not specify how the images are partitioned into training and test sets. Because the constituent crack datasets are collected as image patches from a small number of scenes, an image-level random split can place near-duplicate patches from the same surface in both training and testing, which would inflate the per-dataset Dice/IoU scores in Table 1. Please specify the exact split protocol, report whether any image-level or scene-level overlap exists between training and test, and state how the five evaluation datasets in Table 1 relate to the CRACK500 and Non-crack portions of the compilation.","section":"Dataset"},{"comment":"The Implementation Details section says 'All models are set 5 times of ensemble in the inference' and that STAPLE is used to fuse different samples, but no single-model or fusion-free results are reported. For deterministic baselines, five forward passes produce identical outputs unless test-time augmentation or other stochasticity is used, so the sentence is ambiguous. If STAPLE fusion is applied only to CrossDiff's stochastic samples, the comparison with baselines in Table 1 is not apples-to-apples. Please report CrossDiff with a single sample, with multiple samples without STAPLE, and with the full protocol, and state explicitly what was done for each baseline.","section":"Implementation Details"},{"comment":"Table 3's column labels are not legible in the provided text, and the row layout makes it impossible to determine which variant corresponds to the Cross Encoder alone and which to the Cross Decoder alone. The sentence 'Cross Encoder and Cross Decoder improve 1.36% and 1.35% in average IoU' cannot be verified from the table as printed. Please relabel the rows unambiguously and report each component separately, jointly, and with increasing complexity, so that the contribution of each module to the final 71.61 IoU is clear.","section":"Ablation Studies (Table 3)"},{"comment":"Equation (3) is a classical graph label-propagation recurrence, not the DDPM forward/reverse process used by the proposed model. The paragraph concluding that 'diffusion segmentation performs better when dealing with particularly slender objects' is an analogy rather than a derivation, and the contribution statement claiming 'formulaic validation' overstates the support. Either remove the proof claim or connect Equation (3) to the actual training objective in Equation (4).","section":"Method, Eq. (3)"}],"minor_comments":[{"comment":"Equation (1) writes p(x0:T-1|xT) = product p(xt-1|xt) without conditioning on the input image I; in a conditional diffusion model the reverse step should be p(xt-1|xt, I).","section":"Method, Eq. (1)"},{"comment":"The notation in Equations (2) and (4) is under-specified: the roles of E_I_t, E_x_t, xe, and xd,t are not defined clearly, and the plus sign in Equation (2) could mean concatenation or addition. Please define each symbol and specify the loss weights alpha and beta and the number of diffusion steps T in the Implementation Details.","section":"Method, Eq. (2) and Eq. (4)"},{"comment":"In Table 2, CrossDiff reports identical Dice and IoU values for thresholds 0.3, 0.5, and 0.7, which suggests the fused output is effectively binary; please explain how thresholding interacts with the STAPLE-fused probabilistic outputs.","section":"Results, Table 2"},{"comment":"The claim of an '8.0%' improvement in both Dice and IoU should be accompanied by the exact averaging rule (weighted by sample numbers) and a per-dataset table of improvements, since the unweighted averages from Table 1 differ from the abstract's single number.","section":"Abstract and Main Results"},{"comment":"There are typos in the ablation paragraph: 'donate' should be 'denote' and 'Eecoder' should be 'Encoder'. Also, the phrase 'as the complexity of the Cross Encoder and Cross Decoder increases, it is difficult for the model to converge' is not supported by any convergence curves or training-loss plots.","section":"Ablation Studies"}],"recommendation":"major_revision","confidential_remarks":"The paper should not be accepted in its current form because the headline SOTA claim is not independently checkable: the dataset split is not described and the inference protocol with 5-times ensembling and STAPLE is ambiguous. These issues are potentially fixable, so a major revision is appropriate rather than rejection. If the authors can provide a clear split, verify no leakage, report single-model CrossDiff numbers, and clarify the ablation axes, the result may be publishable. I would also ask the editor to have the novelty claim checked against recent diffusion-based segmentation work, since the related-work section cites only a few diffusion segmentation papers and the 'first for crack segmentation' claim needs a more thorough search."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: the headline result — an 8-point average gain over SegDecNet++ across five crack datasets — is not independently checkable as written. The paper never says how the combined 7169-image dataset is split into train and test, and the five reported test sets could easily overlap with training images because crack datasets are often patches of a handful of scenes. On top of that, CrossDiff is evaluated with a 5-times ensemble and STAPLE fusion while the baselines appear to be single-model. Those two issues together mean the main empirical claim is shaky until the protocol is clarified.\n\nWhat is genuinely new: CrossDiff applies a diffusion probabilistic model to crack segmentation, using a cross-encoder (ViT-style) and cross-decoder with a fusion nexus that injects diffusion timestep information. That specific combination is not in the cited diffusion-segmentation papers (SegDiff, MedSegDiff, DMOISE). The ablation shows the cross modules add about 2.7 IoU on average, which is modest but consistent across datasets. The paper also reports per-dataset numbers on five public benchmarks, which is useful for comparison.\n\nSoft spots, in order of severity: (1) The train/test split is unspecified. If the model was trained on all 7169 images and tested on subsets of the same images, leakage could explain a large part of the gain. The authors must state exactly which images are in each partition and show no overlap. (2) Ensembling. The Implementation Details say \"all models are set 5 times of ensemble,\" but deterministic baselines give identical outputs across runs, so either STAPLE is applied only to the stochastic diffusion model or the baselines are not actually ensembled. A single-sample CrossDiff result is needed, and baselines must match the protocol. (3) The theoretical justification for why diffusion helps slender structures is a qualitative graph-propagation analogy, not a derivation; calling it \"formulaic validation\" overstates it. (4) Equation (4) contains undefined symbols and typos, and the loss is not written in standard DDPM form. (5) No code or error bars.\n\nWho it is for: researchers working on crack/damage segmentation or on applying diffusion models to structured prediction. It could become a useful benchmark reference if the evaluation is cleaned up. I would send it to peer review rather than desk reject, because the architecture is plausible and the questions are answerable in a revision. The referee should push hard for a precise data split, single-model numbers, matched baselines, and a corrected derivation. As it stands, I would not cite the 8% claim.","headline":"The 8% SOTA claim is not yet trustworthy because the evaluation protocol mixes ensembling with an unspecified train/test split; the diffusion-plus-cross-attention idea is plausible but the paper's own ablation limits the claimed contribution.","tokens_in":11513,"tokens_out":3428,"would_cite":false,"duration_ms":35050,"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 introduces CrossDiff, a diffusion probabilistic model with cross-conditional encoder and decoder that, the authors claim, is the first diffusion-based framework for crack segmentation and that outperforms prior state-of-the-art…","keywords":["crack segmentation","diffusion probabilistic model","cross-conditional encoder-decoder","slender crack detection","concrete surface inspection","image segmentation","deep learning"],"falsifier":"Retrain every baseline and CrossDiff under a single protocol: one fixed, published split of the combined dataset, no ensembling, and a single forward pass at inference. If the average Dice/IoU gap between CrossDiff and SegDecNet++ drops below statistical significance (e.g., less than 2–3 points after multiple seeds), the central claim of architecture-driven superiority is refuted.","tokens_in":10442,"feed_emoji":"🧱","tokens_out":4604,"duration_ms":44530,"temperature":0.7,"pith_summary":"The paper proposes CrossDiff, a diffusion probabilistic model adapted to pixel-level crack segmentation, and claims it is the first such model for this task. The authors argue that the iterative denoising process of diffusion models is naturally suited to slender, low-contrast cracks because information propagation along similar pixels keeps thin crack regions coherent. They report state-of-the-art Dice and IoU on five challenging datasets—CFD, CrackTree200, DeepCrack, GAPs384, and Rissbilder—with an average improvement of 8.0 percentage points over the previous best method, with especially large gains on datasets containing the thinnest cracks. If the result holds under equal evaluation conditions, it would make diffusion conditioning a leading approach for fine-structure segmentation in industrial inspection.","feed_headline":"First diffusion model for crack segmentation beats SOTA by 8%","feed_subtitle":"CrossDiff's cross-encoder and cross-decoder keep slender, low-contrast cracks visible across five datasets.","key_machinery":"The central object is a Diffusion Probabilistic Model (DPM) whose step-estimation function is cross-conditioned by an image encoder and a decoder. The forward process adds Gaussian noise to the ground-truth segmentation mask; the reverse process learns to denoise using the raw image as a conditioning signal, written as $\\epsilon(x_t, I, t) = D((E_{I_t} + E_{x_t}, t), t)$, where $E_{I_t}$ is the Cross Encoder embedding of the raw image and $E_{x_t}$ is the embedding of the current noisy mask. The Cross Encoder (patch embedding, positional encoding, Transformer blocks) pulls image priors into the diffusion encoder, the Cross Decoder refines semantic features and adds a direct reconstruction loss, and the Fusion Nexus modulates the fusion with the current noise-level step $t$. This machinery is what the paper claims lets the model keep slender crack structures coherent during iterative denoising.","core_discovery":"CrossDiff couples a standard denoising diffusion backbone (a UNet that progressively denoises a noisy segmentation mask) with a cross-conditional encoder-decoder: a ViT-style Cross Encoder that injects raw-image features into each denoising step, and a Cross Decoder that reconstructs the segmentation mask directly and whose loss also supervises the diffusion encoder. A Fusion Nexus module combines the two feature streams using time-step information. The paper's central claim is that this cross-shaped conditioning lets the model retain crack detail and semantic context simultaneously, yielding better Dice/IoU than CNN- and Transformer-based segmentation baselines. On CFD, for example, it reports 91.34 Dice and 85.54 IoU versus 77.80 and 64.14 for SegDecNet++. The Cross Decoder is used only during training, acting as a regularizer, while inference runs the diffusion chain five times and fuses the samples with the STAPLE algorithm.","pith_inferences":["The 8.0% average improvement is heavily weighted by two thin-crack datasets, so the advantage is plausibly tied to slender-object geometry; the same architecture may transfer to blood-vessel or hair segmentation in medical imaging, which share that geometry.","The paper does not report baseline results with the same 5-time ensemble and STAPLE fusion; a straightforward test is whether the gap persists when every method receives identical inference-time ensembling, since STAPLE alone can boost scores on thin structures.","Since the Cross Decoder is dropped at inference, the inference cost is essentially a standard diffusion UNet plus cross-encoder; ablation could show whether the training-only decoder is necessary for the reported gains or is removable for a cheaper deployment."],"forward_implications":["CrossDiff's reported gains are largest on CFD and CrackTree200, the two datasets with the slenderest cracks, suggesting the method is particularly effective where thin structures make standard segmentation fail.","The training-only Cross Decoder loss acts as a form of deep supervision: the diffusion encoder is optimized both for denoising and for direct segmentation, which may explain the improved Dice and IoU.","Because the method is trained on a combined dataset drawn from seven published crack sources, the same cross-conditional diffusion design should transfer to other industrial defect segmentation tasks, such as corrosion or scratch detection.","The threshold-stability result (Dice/IoU nearly constant between thresholds 0.3 and 0.9) indicates the model produces confident, well-calibrated masks, reducing the need for careful threshold tuning in practice."],"supporting_citations":[{"why":"Supplies the denoising diffusion backbone, the forward/reverse process formulation, and the base training objective that CrossDiff modifies.","marker":"(Ho, Jain, and Abbeel 2020)"},{"why":"Provides improved denoising diffusion details and the noise schedule the CrossDiff implementation follows.","marker":"(Nichol and Dhariwal 2021)"},{"why":"Introduces SegDiff, the diffusion-based segmentation baseline whose image-conditioning approach CrossDiff extends.","marker":"(Amit et al. 2021)"},{"why":"Introduces MedSegDiff, the medical diffusion segmentation work that CrossDiff names as prior art and adapts to crack segmentation.","marker":"(Wu et al. 2024)"},{"why":"Contributes the combined crack dataset (CFD, CRACK500, CrackTree200, DeepCrack, GAPs384, Rissbilder, Noncrack) used for training and evaluation, and the SCCDNet-D32 baseline.","marker":"(Li et al. 2021)"},{"why":"Defines the STAPLE algorithm used to fuse the five inference-time samples of CrossDiff.","marker":"(Warfield, Zou, and Wells 2004)"},{"why":"Presents SegDecNet++, the previous state-of-the-art baseline that CrossDiff claims to beat by 8.0% in average Dice and IoU.","marker":"(Tabernik, ˇSuc, and Skoˇcaj 2023)"}],"fun_headline_variants":["First diffusion model for crack segmentation beats SOTA by 8%","CrossDiff: diffusion-based crack segmentation with 8% SOTA gain","Diffusion model for slender cracks improves segmentation by 8%","CrossDiff: cross-conditional diffusion model beats SOTA by 8%","Crack segmentation: CrossDiff diffusion model exceeds SOTA by 8%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 8-point gain assumes the comparison is fair, yet the paper applies a 5-time ensemble with STAPLE fusion only to CrossDiff and does not describe how the combined dataset was split into training and test sets.","fun_headline_variants_meta":{"raw":{"variants":["First diffusion model for crack segmentation beats SOTA by 8%","CrossDiff: diffusion-based crack segmentation with 8% SOTA gain","Diffusion model for slender cracks improves segmentation by 8%","CrossDiff: cross-conditional diffusion model beats SOTA by 8%","Crack segmentation: CrossDiff diffusion model exceeds SOTA by 8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001319,"raw_usage":{"total_tokens":5368,"prompt_tokens":936,"completion_tokens":4432,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":4337}},"tokens_in":552,"tokens_out":4432,"duration_ms":29984,"temperature":1.0,"reasoning_tokens":4337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:41:38.309610+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain every baseline and CrossDiff under a single protocol: one fixed, published split of the combined dataset, no ensembling, and a single forward pass at inference. If the average Dice/IoU gap between CrossDiff and SegDecNet++ drops below statistical significance (e.g., less than 2–3 points after multiple seeds), the central claim of architecture-driven superiority is refuted.","supporting_citations":[{"cited_title":"K.; Zou, K","cited_arxiv_id":null,"evidence_quote":"Defines the STAPLE algorithm used to fuse the five inference-time samples of CrossDiff."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents SegDecNet++, the previous state-of-the-art baseline that CrossDiff claims to beat by 8.0% in average Dice and IoU."}],"review_version":1}