{"id":"8a9d6e59-d3a1-4358-ab66-01699720a6cb","arxiv_id":"2501.02816","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"InpDiffusion treats image inpainting localization as a conditional mask-generation task with a diffusion model, using edge supervision to refine boundaries, and reports state-of-the-art AUC on Inpaint32K, DID, AutoSplice, IMD, and NIST.","lead":"This paper applies conditional diffusion models to image inpainting localization, predicting which pixels were inpainted by generating masks iteratively with semantic and edge guidance. It reports state-of-the-art AUC scores on several forensics benchmarks and may offer a new generative alternative to standard classification-based tampering detectors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inpaint32K evaluation protocol: no explicit train/test split is stated, so the reported AUCs and the DID generalization numbers may be training-fit scores; this must be settled before the SOTA claim can be accepted.","rationale":"The paper proposes a plausible and novel application of conditional diffusion to image inpainting localization, with a reasonably detailed architecture and an edge-supervision mechanism that is internally coherent. The central claim, however, is empirical: InpDiffusion significantly outperforms existing SOTA methods. That claim stands or falls on the evaluation protocol. The weakest point I find is not the method's internal logic but the Inpaint32K evaluation protocol. The text explicitly says the model is pre-trained on all 32K Inpaint32K images before the DID evaluation, and only specifies 9:1 splits for 'the other datasets.' This leaves open the possibility that Table 1 reports training-set performance. That would explain the unusually large margins over baselines and would invalidate the DID generalization claim, since pretraining on the evaluation domain is a form of leakage. The reader's weakest assumption correctly identified benchmark fairness and missing error bars as concerns, but the train/test ambiguity is more load-bearing because it can invalidate the numbers even if the benchmark is perfectly representative. I do not treat this as evidence of bad faith; it is a concrete, checkable reporting gap. The reader's CONDITIONAL verdict is therefore the right one: the paper should not be accepted as establishing SOTA until the split, code, and a reproduced comparison are provided. My read does not move the verdict, so I leave it unchanged.","tokens_in":13964,"tokens_out":5667,"duration_ms":57238,"concrete_test":"Ask the authors for the exact Inpaint32K train/test split used to produce Table 1, along with code and trained weights. Reproduce Table 1 by training InpDiffusion and the six baselines on the stated training split and evaluating on the stated test split. If no split exists, create a fixed random 90/10 split, run three seeds, and compare the mean and per-seed spread of AUC against ECNet; if the margin shrinks materially below the reported 7.3-point TM gap or overlaps within one standard deviation, the SOTA claim is unsupported. Also confirm that the SNR-shift and loss-ratio ablations in Figures A1-A4 were performed on a validation split disjoint from the test split used in Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim requires that the Table 1 and Table 3 AUCs are generalization numbers. The Experiments section says the model is pre-trained on all 32K Inpaint32K images before evaluating on DID, and only states that 'the other datasets are divided into training and test sets with a 9:1 ratio.' It never specifies the Inpaint32K train/test split. If the 32K images used for pretraining are the same images on which Table 1 reports AUC, then the reported 93.7/96.2/95.6/84.3 are training-fit scores rather than held-out estimates, making the margins over ECNet uninterpretable. The DID evaluation would also be contaminated because the pretraining set includes the evaluation domain. Even if a split exists, the Implementation Details show hyperparameters (T=10, SNR shift -2log(6), lambda:mu=7:3) were selected by performance on Inpaint32K; unless that selection was done on a separate validation split, the reported test AUCs are optimistically biased. This is the load-bearing condition because every downstream claim, including SOTA, generalization, and robustness, inherits from these numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InpDiffusion, a conditional diffusion framework for image inpainting localization (IIL). The method uses an Adaptive Conditional Network (ACN) with a Dual-stream Multi-scale Feature Extractor (DMFE) to obtain semantic and edge conditions from the inpainted image, and a denoising network that iteratively refines the predicted mask while also predicting the tampered edge under Dice loss supervision. The authors report AUC comparisons against six baselines on Inpaint32K, DID, AutoSplice, IMD, and Nist datasets, including ablations of DMFE and edge supervision, and claim that InpDiffusion significantly outperforms existing state-of-the-art methods while improving generalization and robustness.","tokens_in":14226,"tokens_out":3813,"duration_ms":37806,"significance":"If the empirical evaluation survives scrutiny, the paper makes a useful contribution: it is, to my knowledge, the first to formulate IIL as conditional mask generation via a diffusion process, and the edge-supervision strategy for constraining stochastic sampling is well motivated. The paper also provides component ablations, robustness experiments, and a computational complexity comparison, which are strengths. However, the primary benchmark (Inpaint32K) is co-authored by a member of this group, and the most directly comparable baseline (ECNet) is from the same group; these facts are not disqualifying by themselves, but they raise the bar for clarity of the evaluation protocol. The main empirical claim currently rests on an underspecified training/test split for Inpaint32K, so the headline numbers must be interpreted with caution until that is resolved.","major_comments":[{"comment":"The paper does not state whether the Inpaint32K experiments in Tables 1 and 3 use a held-out test split. The only explicit split mentioned is 'the other datasets are divided into training and test sets with a 9:1 ratio,' and the paragraph also says the authors 'pre-train our model using 32K inpainted images from the Inpaint32K dataset and evaluate it on the DID dataset.' This wording suggests that all 32K Inpaint32K images may have been used for training, which would make the Table 1 and Table 3 AUC numbers training-fit scores rather than generalization estimates. The authors must specify the exact train/validation/test partition used for Inpaint32K, including how many images of each inpainting type were held out, and confirm that Table 1 reports evaluation only on the held-out portion.","section":"Experiments (Inpaint32K protocol)"},{"comment":"The hyperparameters T=10, SNR Shift=-2log(6), and the lambda:mu=7:3 ratio are selected based on performance on Inpaint32K, as shown in Supplementary Figures A2-A4. If this selection was performed on the same test set used for the AUC numbers in Table 1, the reported results are optimistically biased by selection on the test set. Please clarify whether a separate validation split of Inpaint32K was used for these ablations, or provide a corrected protocol in which hyperparameters are selected on a distinct validation partition and the test set is used only once.","section":"Implementation Details / Supplementary (hyperparameter selection)"},{"comment":"The statement that 'the other baseline models are also retrained using the Inpaint32K' is not supported by concrete training details. No information is given about training epochs, learning rates, input resolutions, or the exact split used for the baselines, and some baselines (e.g., MVSSNet) were originally pretrained on larger datasets. Without a precise and uniform retraining protocol, it is difficult to verify that the six baselines were evaluated under comparable conditions. Please provide a per-baseline description of the retraining setup and any pretrained-model initialization, and report the resulting performance with the same evaluation split as the proposed method.","section":"Table 1 (baseline retraining)"},{"comment":"Throughout the paper the authors use the word 'significantly' (e.g., 'significantly outperforms,' 'significant superiority'), but no error bars, standard deviations, or statistical tests are reported for any of the AUC values. Since the claimed gains are point estimates from a stochastic training procedure, the authors should either provide confidence intervals or at least repeated-run results, or soften the significance language to avoid implying statistical significance that is not demonstrated.","section":"Experiments (statistical significance)"}],"minor_comments":[{"comment":"The notation 'lambda LWBCE + WIoU (x0_hat, x0)' is ambiguous; it should be written as lambda * (LWBCE(x0_hat, x0) + WIoU(x0_hat, x0)) + mu * LDICE(e_hat, e).","section":"Loss Function, Eq. (8)"},{"comment":"Equations (5) and (6) contain missing parentheses and inconsistent placement of the branch index k, which makes the branch definitions hard to follow; please rewrite them with explicit function arguments.","section":"Method, Eq. (5)-(6)"},{"comment":"The paper claims that InpDiffusion mitigates overconfidence, but no experiment directly measures confidence calibration or overconfidence (e.g., expected calibration error or reliability diagrams). Adding such an evaluation would substantiate this claim.","section":"Abstract / Introduction"},{"comment":"There are several typographical issues, including 'avoide' in the Introduction and 'IpnDiffusion' in the Loss Function section; a careful proofread is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the evaluation protocol for the primary benchmark, Inpaint32K, which is the basis for the strongest SOTA claim. That the dataset and the strongest baseline (ECNet) come from the authors' own group compounds the need for an explicit, arm's-length evaluation protocol. If the authors can clarify the split and confirm that all reported Inpaint32K numbers are held-out, and if they provide retraining details for the baselines, the paper would be publishable. External datasets (DID, AutoSplice, IMD, Nist) do provide some independent evidence, but the Inpaint32K results are the core of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely interesting method for IIL, and the gains over baselines are large. But the paper's central claim rests on numbers from Inpaint32K, and the paper never states whether those are held-out or training-fit. You need to make the authors clarify this before you trust Table 1.\n\nWhat's new: treating IIL as a conditional diffusion mask-generation task, with edge supervision inside the denoising network, is a real twist. The DMFE module is a reasonable multi-scale feature extractor, borrowing from Res2Net and Inception but used sensibly for semantic and edge conditioning. The writing is clear and the method is well motivated. If the empirical results are valid, this is a solid contribution.\n\nWhere it gets soft: the experimental section says DID is exclusively a test set and 'the other datasets are divided into training and test sets with a 9:1 ratio.' Inpaint32K is not described as having such a split. Instead, the authors say they 'pre-train our model using 32K inpainted images' before evaluating on DID. That reads like all 32K images are used for training, which would make the Table 1 AUCs training performance. The ablation in Table 3 and the robustness results in Figure 6 come from the same dataset, so the ambiguity poisons them too. This is not a minor omission; it is the load-bearing condition for the SOTA claim.\n\nThere are lesser issues. No error bars or significance tests, and the hyperparameters (T, SNR shift, loss ratio) were selected by looking at performance on Inpaint32K, presumably on the same images used for reporting. The benchmark Inpaint32K was created by a co-author, and the strongest baseline, ECNet, is from the same group; that alone isn't disqualifying, but combined with the missing split it makes the evaluation look less independent than it should. No code is released.\n\nOn balance, the method deserves a serious referee, but the paper should not be accepted until the authors provide the exact Inpaint32K split, ideally with error bars and code. If the split exists and is simply under-described, this is likely a strong paper. If it doesn't, the SOTA claim is meaningless. A referee should push hard on this specific point.","headline":"Genuinely new diffusion-based IIL method with large reported gains, but the main benchmark's train/test split is unspecified, leaving the SOTA claim uninterpretable until clarified.","tokens_in":14714,"tokens_out":3370,"would_cite":false,"duration_ms":30582,"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":"InpDiffusion reframes image inpainting localization as a conditional mask generation task, using diffusion with semantic and edge guidance to outperform existing methods.","keywords":["image inpainting localization","conditional diffusion models","mask generation","edge supervision","multi-scale feature extraction","image forensics","AI-generated image detection"],"falsifier":"Re-running the Inpaint32K comparison with multiple random seeds and reporting confidence intervals, or evaluating InpDiffusion on an independently constructed inpainting dataset from a different research group, would settle the claim; if the roughly 7-point AUC gap over ECNet shrinks or reverses, the state-of-the-art claim collapses.","tokens_in":13784,"feed_emoji":"🎯","tokens_out":8207,"duration_ms":62909,"temperature":0.7,"pith_summary":"Image inpainting localization (IIL) is usually treated as a pixel classification problem, and the paper argues this makes models overconfident and blind to subtle tampering boundaries. The proposed method, InpDiffusion, instead treats IIL as a conditional mask generation task inside a diffusion model: the denoising process is guided by semantic and edge features extracted from the image, and an edge-supervision loss keeps the stochastic sampling from eroding fine boundaries. The authors claim that this generative formulation, with the Dual-stream Multi-scale Feature Extractor (DMFE), significantly outperforms six state-of-the-art baselines; for example, on the Inpaint32K dataset it reports AUC of 93.7% versus 86.4% for the best baseline on traditional-method inpaintings. A sympathetic reader would care because the method points to a new family of generative, condition-guided detectors for AI-manipulated images, with reported gains across CNN-, GAN-, and diffusion-based inpaintings as well as on real-life datasets.","feed_headline":"Diffusion model localizes inpainted regions with edge guidance","feed_subtitle":"Framing inpainting localization as mask generation lifts AUC from 86.4 to 93.7 on Inpaint32K.","key_machinery":"The machinery that carries the argument is the conditional denoising network together with its edge-supervision loop. The Adaptive Conditional Network (ACN), built on a Pyramid Vision Transformer backbone, extracts hierarchical features from the concatenation of the image and the current noisy mask, with the diffusion time step injected as a token; its Dual-stream Multi-scale Feature Extractor (DMFE) uses two streams of dilated convolutions with rates $\\{3,5,7\\}$ in increasing and decreasing order to produce semantic and edge conditions. The Denoising Network (DN) is a U-shaped encoder–decoder with adaptive group normalization, producing both the denoised mask $\\hat{x}_0$ and the denoised edge $\\hat{e}$ at every step; the loss $L_{\\text{total}} = \\lambda L_{\\text{WBCE+WIoU}}(\\hat{x}_0, x_0) + \\mu L_{\\text{Dice}}(\\hat{e}, e)$ balances mask fidelity against edge fidelity. This edge supervision is what allows the model to keep stochastic sampling (which avoids overconfident point estimates) without losing subtle tampering boundaries.","core_discovery":"The paper's central claim is that image inpainting localization is best posed as a generative, conditional mask-generation problem rather than as a discriminative segmentation problem. InpDiffusion trains a denoising network to predict the clean tampering mask $\\hat{x}_0$ from a noisy mask, conditioning each denoising step on image semantic features and on edge features extracted by an Adaptive Conditional Network (ACN). A second decoder predicts the inpainted object's edge $\\hat{e}$, and a Dice loss supervises this edge prediction, balancing the mask-supervision loss (weighted BCE plus weighted IoU) so that excessive randomness in sampling does not destroy subtle boundaries or cause overconfident mispredictions. The paper further introduces the Dual-stream Multi-scale Feature Extractor (DMFE), a two-stream dilated-convolution module with increasing and decreasing dilation rates, to enrich the semantic and edge conditions. Extensive experiments report state-of-the-art AUC across Inpaint32K, DID, AutoSplice, IMD, and NIST, with ablations showing that both DMFE and edge supervision each contribute to the gains.","pith_inferences":["Because the diffusion sampler can be run multiple times, the architecture naturally supports uncertainty estimates over the predicted mask; the paper does not report such estimates, but they would be a straightforward extension.","A direct test of the method's forensic value would be to apply it to images edited by newer generative models not represented in Inpaint32K; the paper does not run this experiment.","The main benchmark was created by one of the paper's co-authors, so an independent, externally built inpainting-localization benchmark would be the most convincing check on the claimed state-of-the-art performance."],"forward_implications":["If the method generalizes as claimed, forensic pipelines can adopt conditional diffusion for a wider class of manipulation-localization tasks, including splicing and copy-move detection.","The edge-supervision strategy offers a general recipe for controlling stochastic sampling in conditional diffusion models applied to pixel-level prediction tasks.","The reported cross-dataset results imply that pre-training on a diverse inpainting corpus like Inpaint32K can transfer to unseen inpainting techniques and real-world manipulated images.","The DMFE design suggests that bidirectional multi-scale dilated convolutions improve conditioning-feature quality for denoising, a component that could be reused in other guided diffusion frameworks."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model (DDPM) forward and reverse processes that InpDiffusion adapts for mask generation.","marker":"Ho, Jain, and Abbeel 2020"},{"why":"Provides the DDIM-style formulation in which the network directly estimates the clean mask $\\hat{x}_0$.","marker":"Song, Meng, and Ermon 2020"},{"why":"Supplies adaptive group normalization used to inject time-step information into the denoising network.","marker":"Dhariwal and Nichol 2021"},{"why":"Provides the SNR-based variance schedule and SNR shift used in training InpDiffusion.","marker":"Hoogeboom, Heek, and Salimans 2023"},{"why":"Supplies the Dice loss used for edge supervision in the denoising network.","marker":"Xie et al. 2020"},{"why":"Supplies the weighted binary cross-entropy and weighted IoU losses used for mask supervision.","marker":"Wei, Wang, and Huang 2020"},{"why":"Provides the Pyramid Vision Transformer backbone used in the Adaptive Conditional Network.","marker":"Wang et al. 2022"},{"why":"Res2Net's multi-scale design inspired the Dual-stream Multi-scale Feature Extractor.","marker":"Gao et al. 2019"},{"why":"Defines the Inpaint32K dataset, the main benchmark on which the state-of-the-art claim is made.","marker":"Hao 2024"},{"why":"ECNet, the strongest baseline that InpDiffusion must beat on Inpaint32K and other datasets.","marker":"Hao et al. 2024"}],"fun_headline_variants":["Diffusion mask generation spots inpainted regions","Edge-supervised diffusion refines inpainting detections","Conditional diffusion beats state-of-the-art in IIL","Mask diffusion with edge cues boosts localization","Generative inpainting localization via denoising network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim of state-of-the-art performance rests on the assumption that the Inpaint32K benchmark is a fair and representative evaluation, that the six baselines were retrained under comparable conditions, and that the reported point-estimate AUC differences are genuine despite the absence of error bars or significance tests.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion mask generation spots inpainted regions","Edge-supervised diffusion refines inpainting detections","Conditional diffusion beats state-of-the-art in IIL","Mask diffusion with edge cues boosts localization","Generative inpainting localization via denoising network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1317,"prompt_tokens":1008,"completion_tokens":309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":624,"tokens_out":309,"duration_ms":3961,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:03:06.194586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the Inpaint32K comparison with multiple random seeds and reporting confidence intervals, or evaluating InpDiffusion on an independently constructed inpainting dataset from a different research group, would settle the claim; if the roughly 7-point AUC gap over ECNet shrinks or reverses, the state-of-the-art claim collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DDIM-style formulation in which the network directly estimates the clean mask $\\hat{x}_0$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Dice loss used for edge supervision in the denoising network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the weighted binary cross-entropy and weighted IoU losses used for mask supervision."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Pyramid Vision Transformer backbone used in the Adaptive Conditional Network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Res2Net's multi-scale design inspired the Dual-stream Multi-scale Feature Extractor."}],"review_version":1}