{"id":"093432dd-4757-4d33-9b18-fad7e1aae1f9","arxiv_id":"2412.12503","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-branch segmentation network that fuses multi-scale RGB and noise features and adds predicted edge masks reports state-of-the-art F1 on three splicing localization benchmarks.","lead":"This paper proposes a deep network for image splicing localization that fuses multi-scale features from the RGB image and its noise fingerprint, then adds predicted edge masks to guide the segmentation. A generalist might read it because it claims state-of-the-art accuracy on three standard forgery benchmarks without releasing code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA margins are likely confounded by 120k-image pre-training and heterogeneous baseline protocols, so the architecture's contribution is not yet established.","rationale":"The paper's novelty is a combination of known modules (SegFormer, NoisePrint++, CondConv, SCCM), so the central contribution is the claimed empirical superiority. That claim is supported only by Table I. The most fragile link is the comparison protocol, because the manuscript explicitly mixes 'from original papers' numbers with 'running source code' numbers, and the authors' own model receives a 120k-image pre-training that baselines may not have. Table II's RGB-only row demonstrates that even without the proposed modules the model beats most baselines, suggesting the pre-training dominates. Other potential issues (e.g., Eq. (4) does not visibly use the edge mask E_i, so the 'edge-supervision' is only in the loss) are secondary: even if the edge-supervision description is incomplete, the empirical margin is the headline claim. Thus a controlled re-run is the single check that would settle whether the SOTA claim is real. This matches the reader's weakest assumption, so no verdict change is needed.","tokens_in":10368,"tokens_out":7115,"duration_ms":62093,"concrete_test":"Run a controlled re-benchmark: use the official code of all six baselines (ManTra-Net, PSCC-Net, MVSS-Net++, HiFi-Net, FARA-Net, D-Net) and the authors' released model, train each with the same pre-training data (or all with pre-training disabled) and the same fine-tune split as [3], then evaluate all with a single script at 256x256. If the F1 gap between the proposed method and the strongest baseline drops below ~1 point on any dataset, the reported SOTA improvement is an artifact of pre-training/protocol rather than the architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section III-B: 'our method achieves the best performance on the three datasets... a large improvement in P, R and F1') rests entirely on Table I. The protocol section says the model is pre-trained on 120k DEFACTO+PSCC images and then fine-tuned on the same split as FARA-Net [3], while baseline scores are 'either taken from their original papers or by running the publicly available source code.' This mixes training protocols: the proposed model benefits from 120k extra pre-training images, and each baseline may have been trained elsewhere, on different splits, resolutions, or preprocessing. The ablation (Table II) makes the confound visible: the RGB-only variant (no noise branch, no CSF/CDF, no edge supervision) already scores F1=92.3 on Columbia and 95.9 on NIST16, exceeding or matching most baselines listed in Table I (e.g., PSCC-Net 81.4 and 96.4, HiFi-Net 82.1 and 96.2). Since the only major difference between that row and the compared baselines is the 120k pre-training, the claimed margin of 3.2-4.1 F1 points cannot be attributed to the proposed CSF/CDF/edge modules without a matched-protocol comparison. No code, test split lists, or per-baseline protocol descriptions are provided, so the reader cannot independently verify whether the comparison is fair.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dual-branch end-to-end network for image splicing localization. An RGB image and a NoisePrint++-derived noise image are each fed into a SegFormer backbone to produce multi-scale features. These are aggregated by a cross-scale fusion (CSF) and a cross-domain fusion (CDF) module. A separate branch predicts an edge mask using Sobel-based edge blocks, and the fused features and edge mask are combined through SCCM attention heads in a progressive localization scheme. The loss combines BCE on four localization masks and Dice on the predicted edge mask (with the edge ground truth obtained by a Sobel projection of the segmentation ground truth). Experiments on CASIAv2, Columbia, and NIST16 report F1 scores of 92.3, 97.1, and 98.7, respectively, which the authors claim outperform state-of-the-art methods. Ablations show monotonic F1 improvements as modules are added.","tokens_in":10664,"tokens_out":4190,"duration_ms":36837,"significance":"If the reported results are reliable, the architecture offers a practical improvement in splicing localization, and the explicit edge-supervision mechanism is a plausible and well-motivated component. The paper is clearly written in its architecture description, uses established components (SegFormer, CondConv, SCCM), and the ablation study is internally monotonic, which is a positive sign. However, the central claim of state-of-the-art superiority rests on Table I, and the comparison protocol there is not sufficiently controlled to support that claim. In particular, the proposed model is pre-trained on 120k images before fine-tuning, while the baselines' results are taken from heterogeneous sources, making the reported margins uninterpretable as architecture-level gains. The paper also lacks error bars, significance tests, and full reproducibility details. These gaps are load-bearing for the main claim.","major_comments":[{"comment":"The state-of-the-art comparison is confounded by unmatched training protocols. The proposed model is pre-trained on 120k DEFACTO and PSCC images (Section III-A1) before fine-tuning on the target datasets, whereas the baseline results are 'either taken from their original papers or by running the publicly available source code' (Section III-B). This means that each baseline may have been trained on different splits, at different resolutions, and with different pre-training. The confound is visible in Table II: the RGB-only variant (no noise branch, no CSF/CDF, no edge supervision) already scores F1=92.3 on Columbia and 95.9 on NIST16, exceeding or matching the majority of baselines listed in Table I (e.g., PSCC-Net 81.4 and 96.4, HiFi-Net 82.1 and 96.2). Because the only major difference between that row and the compared baselines is the 120k pre-training, the claimed improvement of 3.2-4.1 F1 points cannot be attributed to the proposed CSF/CDF/edge modules. A matched-protocol comparison, where all baselines are fine-tuned from the same pre-training and evaluated on the same split, is required to support the superiority claim.","section":"Section III-B, Table I"},{"comment":"The ablation results are presented without error bars, confidence intervals, or significance tests, and appear to come from a single run. The text in Section III-C reports that CSF improves performance by 'approximately 1.2%', CDF by 'about 0.5%', and edge supervision by 'around 1.7%', but several of these increments are small (e.g., CASIAv2 F1: 90.7 to 91.2 after adding CDF) and could be within run-to-run variance, especially given a batch size of 10 and only 25 epochs. The authors should either run multiple seeds and report mean and standard deviation, or provide significance tests, before drawing quantitative conclusions about the per-module contribution.","section":"Section III-C, Table II"},{"comment":"The reproducibility of the study is insufficient for the core claim. The training/testing split is specified only by reference to [3] ('we follow the same training/testing split on COLUMBIA, CASIAv2 and NIST16 datasets, as in [3]'), but the actual split (e.g., which images are in train versus test, and any class balancing) is not described. No code, split lists, or test-time preprocessing details (e.g., whether test images are resized to 256×256 and whether post-processing such as CRF is applied) are provided. Given that the comparison with prior work depends on identical evaluation conditions, this lack of detail prevents independent verification of the reported margins.","section":"Section III-A, Section III-B"},{"comment":"The robustness claim is not substantiated quantitatively. The text states that the method 'achieves better performance on attacked images' after applying Resize (ratio 0.9) and Gaussian Noise (variance 3), and that this 'can be clearly seen from Fig. 2', but no numerical results, evaluation protocol for the attacked images, or error bars are provided. The figure appears to be a qualitative plot without axes labels or legend clarity. This claim should either be removed or supported with a table of metrics under the stated attacks.","section":"Section III-B, Fig. 2"}],"minor_comments":[{"comment":"Equation (3) defines Ei = Conv(EB(Conv(Ri ⊕ Ei−1))), but the surrounding text describes the Edge Block as containing a Sobel operator and a convolutional layer, and describes a multiplication of the edge prior with the RGB features. The equation does not match this description; please clarify the exact operations inside EB and how the multiplication with Ri is represented.","section":"Section II-C, Eq. (3)"},{"comment":"Equation (1) uses the notation (Ri−1; Ri; Ri+1) for the cross-scale fusion, but the text immediately states that for i = 1 only R1 and R2 are fused, and for i = 4 only R3 and R4 are fused. The boundary cases should be expressed explicitly in the equation or in a footnote.","section":"Section II-B, Eq. (1)"},{"comment":"Table I contains an extraneous run of Unicode-like garbage characters after the 'Ours' row (strings such as '/uni00000030/uni00000039/...'). This is clearly a formatting corruption and must be removed before publication.","section":"Table I"},{"comment":"The phrase 'slight worse' should be 'slightly worse'.","section":"Section III-B"},{"comment":"The loss in Eq. (5) applies BCE to all four localization masks and Dice only to the edge mask E4. It would be clearer if the authors stated whether the edge ground truth GE is derived from G4 (the lowest-resolution ground truth) and whether the Dice term is computed at the same resolution as E4.","section":"Section II-E"}],"recommendation":"major_revision","confidential_remarks":"The core architectural contribution is plausible and the ablation trend is encouraging, but the state-of-the-art claim is not currently established because of the unmatched comparison protocol and the absence of statistical validation. I recommend major revision: the authors should re-run all baselines under the same pre-training and fine-tuning protocol, provide error bars, and fully specify the data split. The editor may also wish to ask the authors to disclose whether any pre-training images (DEFACTO/PSCC) overlap with the evaluation datasets, since this could further confound the comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The architecture is a plausible assembly of known components into a dual-branch splicing localization network, and the paper's own ablation story is coherent. But the headline 3-4 point F1 gain over SOTA rests on a comparison that is not matched: the model gets 120k extra pre-training images, while baselines come from their original papers or unofficial runs.\n\nWhat is actually new: the specific combination of SegFormer with NoisePrint++ input, cross-scale then cross-domain fusion via CondConv, and a learned edge mask supervised by a Sobel projection of the same ground truth. I don't see a derivation fitted to the benchmark result, and the edge supervision is not circular, since it is trained on a transformed mask, not the test metric. The internal ablation (Table II) is monotonic: each added module improves F1 by about 1-2 points, which is exactly the kind of trend you want to see.\n\nThe real soft spot is the external comparison. The ablation row 'RGB'—their backbone with none of the proposed modules—already scores 92.3 F1 on Columbia and 95.9 on NIST16, higher than PSCC-Net, HiFi-Net, MVSS-Net++, and within a point of D-Net on Columbia. Since the only major difference between that row and those baselines is the 120k pre-training, the margin cannot be attributed to the CSF/CDF/edge modules without re-training baselines under the same protocol. The paper also gives no error bars, no significance tests, and the split is only described by reference to [3]. No code is promised, which makes the central claim unverifiable.\n\nThat said, I don't want to overstate the problem. The method is a legitimate extension of an ongoing research program, the writing is clear, and the robustness experiments suggest the full model degrades gracefully. If the authors re-run baselines with the same pre-training and release code, the numbers could hold up. As it stands, the paper is worth reviewing because the research community would benefit from a matched evaluation of these components. I'd accept it for peer review but would insist on a uniform protocol and an explicit ablation of pre-training contribution. I wouldn't cite the SOTA numbers in my own work until that happens.","headline":"A plausible architecture with an internally coherent ablation, but the headline SOTA margin is not established because the comparison is mismatched: 120k extra pre-training for the proposed model versus baselines taken from original papers.","tokens_in":11175,"tokens_out":2210,"would_cite":false,"duration_ms":19852,"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":"Explicit edge supervision plus multi-scale RGB-noise fusion lifts splicing localization to state-of-the-art F1 scores on three public benchmarks.","keywords":["image splicing localization","multi-scale features","cross-scale fusion","cross-domain fusion","edge supervision","noise fingerprint","image forensics"],"falsifier":"Rerun the six baselines and the proposed network under identical training data, input size, and post-processing on CASIAv2, Columbia, and NIST16, and check whether the reported F1 margins (3.2, 4.1, and 2.3 points) survive; if the gaps vanish or invert when all methods share the same protocol, the central claim is not attributable to the architecture.","tokens_in":10195,"feed_emoji":"🔍","tokens_out":7335,"duration_ms":57308,"temperature":0.7,"pith_summary":"The paper sets out to show that image splicing can be localized more accurately when the network is explicitly taught the boundary of the forged region while also merging multi-scale features from two complementary views: the RGB image and its sensor noise fingerprint. The authors build an end-to-end dual-branch network with a SegFormer backbone, a cross-scale fusion step that combines neighboring feature resolutions, and a cross-domain fusion step that mixes RGB and noise features with conditionally parameterized convolutions. A separate edge-mask branch predicts the forged-region boundary from Sobel-based edge blocks, and that edge mask is folded back into the localization heads through attention to supervise training incrementally. The paper reports state-of-the-art F1 scores on three public benchmarks, 92.3 on CASIAv2, 97.1 on Columbia, and 98.7 on NIST16, with gains of 2.3 to 4.1 F1 points over the best listed baseline. A reader should care because better edge exploitation is a path to reducing both missed regions and false alarms in practical image forensics.","feed_headline":"Edge-supervised fusion network sets new splicing localization records","feed_subtitle":"A dual-branch RGB-plus-noise network with edge supervision gains 2.3 to 4.1 F1 points over previous best methods on three benchmarks.","key_machinery":"The carrying mechanism is a three-stage dual-branch architecture. Cross-Scale Fusion resamples and concatenates neighboring resolution layers from the SegFormer backbone so that both local boundary detail and global inconsistency are present. Cross-Domain Fusion uses conditional convolution (CondConv) to adaptively combine RGB features with noise-fingerprint features, exploiting the complementarity between visible appearance and sensor-level traces. Edge Mask Prediction builds Sobel-operator edge blocks in a progressive chain, multiplying the predicted edge prior back into the RGB features so boundary artifacts are highlighted. Edge-Supervision Localization then runs four SCCM heads (spatial plus channel attention) that consume the fused features together with the reshaped edge mask, producing progressively refined masks; the final loss is a sum of BCE losses on the four forgery masks plus a Dice loss on the predicted edge mask, with the edge ground truth obtained by a Sobel operation on the coarsest mask.","core_discovery":"The central claim is that edge information, treated as a first-class supervision signal rather than a byproduct, is what pushes splicing localization past existing methods. The network takes an RGB image and its NoisePrint++ noise image, runs each through a SegFormer encoder to get four scales of features, fuses neighboring scales, then fuses the RGB and noise domains via CondConv. From the RGB features it predicts a boundary mask using stacked Sobel-based edge blocks, and the predicted edge mask is multiplied back into the fused features before four SCCM attention heads produce progressively refined forgery masks. On the three standard datasets the authors report the best published F1 in their comparison, with precision, recall, and F1 all improved; for instance, against D-Net on Columbia the F1 gain is 4.1 points, and against PSCC-Net on CASIAv2 it is 3.2 points. The ablation study attributes the gain to the combination of cross-scale fusion, cross-domain fusion, and edge supervision, with edge supervision contributing the single largest increment.","pith_inferences":["A natural stress test would blur or compress the splices' seams: if the reported margin comes from edge supervision, sharp boundaries are the fuel, and seam-smoothing should shrink the F1 gap.","Because the cross-domain branch depends on a sensor-noise fingerprint, images that are heavily denoised or upsampled may erode the advantage; the paper's robustness experiments cover resize and Gaussian noise but not re-compression, so JPEG re-encoding remains an untested boundary.","The 120k-image pre-training on DEFACTO and PSCC is large relative to the fine-tuning sets, so a controlled comparison with smaller or no pre-training would separate the architecture's contribution from the scale of pre-training."],"forward_implications":["Explicitly supervising the forged boundary can reduce false alarms and improve region integrity, since the edge mask concentrates training on the most decisive pixels.","Combining RGB and noise-fingerprint domains in a transformer segmentation backbone yields complementary cues that improve pixel-level localization over either domain alone.","On the three standard benchmarks (CASIAv2, Columbia, NIST16), the method reports F1 gains of 3.2, 4.1, and 2.3 points over the best listed baselines, suggesting practical headroom for forensic tools.","The progressive multi-scale supervision makes the localization robust to differences in forged-region scale, including small spliced patches."],"supporting_citations":[{"why":"PSCC-Net serves as a baseline in Table I and contributes the SCCM localization head used in the edge-supervision stage.","marker":"[15]"},{"why":"D-Net is the strongest baseline on Columbia; the paper measures its 4.1-point F1 margin against it.","marker":"[2]"},{"why":"MVSS-Net++ is both a baseline and the inspiration for extracting edge artifacts from multi-scale features.","marker":"[19]"},{"why":"FARA-Net supplies the training/testing split that the method adopts on the three benchmark datasets.","marker":"[3]"},{"why":"NoisePrint++ converts the RGB input into the noise image that forms the second branch.","marker":"[8]"},{"why":"SegFormer is the backbone that supplies the four multi-scale RGB and noise feature maps.","marker":"[17]"},{"why":"CondConv implements the adaptive kernel in the cross-domain fusion stage.","marker":"[18]"},{"why":"Columbia is one of the three evaluation datasets and the one where the method reports its largest F1 gain.","marker":"[24]"},{"why":"CASIAv2 is the evaluation dataset with the hardest forged regions, where the method reports F1 of 92.3.","marker":"[25]"},{"why":"NIST16 is the third evaluation dataset, where the method reports F1 of 98.7.","marker":"[26]"}],"fun_headline_variants":["Edge supervision boosts splicing localization F1 by up to 4.1 points","Multi-scale fusion with edge masks beats state-of-the-art on three benchmarks","New network combines RGB and noise features for sharper splicing detection","Edge-aware cross-fusion wins on Columbia, CASIAv2, and more"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole superiority claim rests on the assumption that the comparison is fair: every baseline was evaluated with the same test split, input resolution, and post-processing, and none of the 120,000 pre-training images overlaps the benchmark test images.","fun_headline_variants_meta":{"raw":{"variants":["Edge supervision boosts splicing localization F1 by up to 4.1 points","Multi-scale fusion with edge masks beats state-of-the-art on three benchmarks","New network combines RGB and noise features for sharper splicing detection","Edge-aware cross-fusion wins on Columbia, CASIAv2, and more"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001152,"raw_usage":{"total_tokens":4775,"prompt_tokens":944,"completion_tokens":3831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":3752}},"tokens_in":560,"tokens_out":3831,"duration_ms":26125,"temperature":1.0,"reasoning_tokens":3752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:00:33.625566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the six baselines and the proposed network under identical training data, input size, and post-processing on CASIAv2, Columbia, and NIST16, and check whether the reported F1 margins (3.2, 4.1, and 2.3 points) survive; if the gaps vanish or invert when all methods share the same protocol, the central claim is not attributable to the architecture.","supporting_citations":[{"cited_title":"PSCC-Net: Progressive Spatio-Channel Correlation Network for Image Manipula- tion Detection and Localization","cited_arxiv_id":null,"evidence_quote":"PSCC-Net serves as a baseline in Table I and contributes the SCCM localization head used in the edge-supervision stage."},{"cited_title":"D-Net: A dual-encoder network for image splicing forgery detection and localization","cited_arxiv_id":null,"evidence_quote":"D-Net is the strongest baseline on Columbia; the paper measures its 4.1-point F1 margin against it."},{"cited_title":"MVSS-Net: Multi-View Multi-Scale Supervised Networks for Image Manipulation Detection","cited_arxiv_id":null,"evidence_quote":"MVSS-Net++ is both a baseline and the inspiration for extracting edge artifacts from multi-scale features."},{"cited_title":"Feature Aggregation and Region-Aware Learning for Detection of Splicing Forgery","cited_arxiv_id":null,"evidence_quote":"FARA-Net supplies the training/testing split that the method adopts on the three benchmark datasets."},{"cited_title":"TruFor: Leveraging All-Round Clues for Trust- worthy Image Forgery Detection and Localization","cited_arxiv_id":null,"evidence_quote":"NoisePrint++ converts the RGB input into the noise image that forms the second branch."},{"cited_title":"Alvarez, and Ping Luo","cited_arxiv_id":null,"evidence_quote":"SegFormer is the backbone that supplies the four multi-scale RGB and noise feature maps."},{"cited_title":"Cond- Conv: Conditionally Parameterized Convolutions for Efficient Inference","cited_arxiv_id":null,"evidence_quote":"CondConv implements the adaptive kernel in the cross-domain fusion stage."},{"cited_title":"Hsu and S.-F","cited_arxiv_id":null,"evidence_quote":"Columbia is one of the three evaluation datasets and the one where the method reports its largest F1 gain."},{"cited_title":"CASIA Image Tampering Detection Evaluation Database","cited_arxiv_id":null,"evidence_quote":"CASIAv2 is the evaluation dataset with the hardest forged regions, where the method reports F1 of 92.3."},{"cited_title":"Yates, Andrew Delgado, Daniel Zhou, Timothee Kheyrkhah, Jeff Smith, and Jonathan Fiscus","cited_arxiv_id":null,"evidence_quote":"NIST16 is the third evaluation dataset, where the method reports F1 of 98.7."}],"review_version":1}