Pith. sign in

REVIEW 4 major objections 7 minor 84 references

M2SFormer: Multi-Spectral and Multi-Scale Attention with Edge-Aware Difficulty Guidance for Image Forgery Localization

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read M2SFormer claims state-of-the-art cross-domain image forgery localization by fusing multi-frequency and multi-scale attention in a Transformer encoder–decoder, reporting best or second-best DSC/mIoU on most of six unseen test datasets.

desk verdict New architecture with a plausible design, but the central SOTA claim is not supported because two comparison tables contain impossible DSC/mIoU pairs; this needs correction before the results can be trusted. read the letter →

arxiv 2506.20922 v1 pith:TG5RDOSD submitted 2025-06-26 cs.CV

Ju-Hyeon Nam , Dong-Hyun Moon , Sang-Chul Lee This is my paper · ORCID
classification cs.CV
keywords imageforgerylocalizationmulti-spectralattentionmulti-scaletransformerencoder-decoderdifficulty-guidedcross-domaingeneralization2DDCTcurvature-baseddifficulty
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a single Transformer-based network can localize manipulated image regions across domains it never trained on, without any fine-tuning on the target dataset. The proposed model, M2SFormer, claims to outperform twelve existing forgery-localization and segmentation baselines on most of six external test datasets under two independent training schemes, while using fewer FLOPs (14.2G) than most competitors. The motivation is practical: forged images arrive from unknown pipelines, so a model that generalizes to unseen manipulation types matters for real-world content verification. The authors argue the gains come from unifying frequency- and spatial-domain attention in the skip connections and from a difficulty-guided decoder that concentrates computation on boundary regions where subtle tampering is hardest to spot.

What carries the argument

The load-bearing component is the M2S (Multi-Spectral with Multi-Scale) attention block, which operates on skip-connection features after they are channel-reduced, resized, and concatenated into one cross-scale map. Multi-spectral attention computes top-K 2D DCT coefficients (basis images $D^{u_k,v_k}_{h,w}$) of that map, pools them via global average and global max, and passes them through a shared $1\times1$ convolution block to produce a channel attention map $M_{\text{spectral}}$; multi-scale attention then downsamples the recalibrated map into a pyramid, applies dilated $3\times3$ convolutions and learnable foreground/background weights $\alpha^l_i, \beta^l_i$ at each level, and sums the levels back with a residual connection. The second mechanism is the Edge-Aware DGA-based Transformer decoder: a global prior map $G$ from the deepest feature yields a curvature map $\kappa_i = (G_x^2 G_{yy} - 2 G_x G_y + G_y^2 G_{xx}) / (G_x^2 + G_y^2)^{1.5}$, edge-weighted averaging gives a difficulty score $s$, and the thresholded text label 'hard'/'easy' is embedded via byte-pair encoding and applied to each decoder feature map as channel-wise attention. In the authors' telling, the M2S block captures both frequency artifacts and scale-varying boundary cues in one pass, while the DGA mechanism steers decoder compute toward high-curvature, edge-dense regions where subtle forgeries hide; the whole network is trained end-to-end with a two-term BCE loss on the final mask and the upsampled global prior map.

What would settle it

Re-run the full comparison twice: once with each baseline using its original published training recipe (including its multi-task losses and input resolution), and once with a per-method hyperparameter search on a held-out validation split of the training dataset. If M2SFormer no longer achieves best or second-best DSC/mIoU on a majority of the six unseen datasets under either protocol, the paper's central generalizability claim would be refuted. A cheaper first check is to train MVSSNet, EITLNet, and PIMNet with their published settings and observe whether any of them matches M2SFormer's unseen-domain numbers.

Watch

Extended reading notes

Core claim

At the paper's core is the claim that spatial and frequency cues should not be processed in separate branches, as in earlier dual-encoder or multi-task designs, but fused inside a single attention block placed in the skip connection of a Transformer encoder–decoder. M2SFormer uses 2D DCT basis images to build a channel-attention map over the cross-scale feature map, then applies SIFT-inspired multi-scale spatial attention with learnable foreground/background balance at each pyramid level. In the decoder, an Edge-Aware Difficulty Calculator derives a curvature map from the deepest feature map, averages curvature only inside edge regions, and thresholds it into a text label ('hard' or 'easy'); that label is tokenized with byte-pair encoding and applied as channel-wise attention to guide each decoder stage. With this machinery, the paper reports the best or second-best DSC and mIoU on most of six unseen datasets (CASIAv1, Columbia, IMD2020, CoMoFoD, In the Wild, MISD) under both CASIAv2 and DIS25k training schemes, asserting superior cross-domain generalization at 27.4M parameters and 14.2G FLOPs.

Load-bearing premise

The central outperformance claim rests on the assumption that training every baseline with the same fixed recipe — batch size 32, learning rate $10^{-4}$, 100 epochs, $256\times256$ resolution — is a fair comparison, even though several baselines were originally published with specialized training schemes, multi-task losses, or higher-resolution inputs; if those settings were not re-tuned for this paper, the reported gains could reflect handicapped competitors rather than the M2SFormer design.

Editorial extensions

If this is right

  • A model trained on one forgery dataset (CASIAv2 or DIS25k) transfers to six unseen datasets, including real-world 'In the Wild' images, with no target-domain fine-tuning — the configuration needed to deploy forgery detection on new data.
  • Fusing frequency attention in feature space rather than at the input image keeps costs low: 14.2G FLOPs and 27.4M parameters, below most Transformer and dual-encoder baselines in the paper's comparison.
  • The M2S block is backbone-agnostic in the paper's experiments: swapping in ResNet50, Res2Net50, ResNeSt50, MiT-B2, or P2T-Small keeps the same architecture and still yields higher unseen-domain DSC than the CNN-type baselines.
  • The edge-aware difficulty label ('hard'/'easy') raises unseen-domain DSC from 32.3 to 43.0 in the paper's ablation, which the authors take as evidence that boundary-focused difficulty guidance, not just multi-scale fusion, drives cross-domain robustness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to transplant the M2S attention block and curvature-based difficulty label into other dense prediction tasks with subtle cues, such as camouflaged object detection or industrial defect localization; the paper mentions general transfer in its broader-impact section but demonstrates nothing outside forgery localization.
  • Because the difficulty score is computed per image and per decoder stage, a natural follow-up the paper does not explore is variable compute: using the 'easy' label to skip decoder stages or lower resolution at test time, turning the mechanism into an adaptive-efficiency device.
  • The choice to apply 2D DCT at the feature level rather than the pixel level implies a design principle — frequency analysis on aggregated features can substitute for input-level transforms — that could be tested against pixel-level DCT methods on the same benchmarks to isolate what the feature-level fusion actually buys.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes M2SFormer, a Transformer encoder-decoder for pixel-level image forgery localization. Its main components are (i) an M2S attention block inserted into the skip connections, which combines 2D-DCT-based multi-spectral channel attention with a SIFT-inspired multi-scale spatial attention pyramid, and (ii) an Edge-Aware Difficulty Calculator that derives a global prior map from the deepest decoder feature, computes a curvature-based difficulty score, converts it into a text label ("hard" or "easy"), and uses that label to modulate channel-wise attention in the decoder. The method is evaluated under two training schemes (CASIAv2 and DIS25k) and then tested on six external datasets, with five-fold cross-validation, and is reported to outperform twelve baselines on most unseen-domain metrics while using 14.2 GFLOPs. The paper also includes ablations of the M2S block and the DGA decoder and a backbone study.

Significance. If the reported results are trustworthy, the paper would offer a practically useful cross-domain forgery localization model with a reasonably efficient architecture, and the experimental design has strengths: five-fold cross-validation, six external unseen-domain datasets, ablations of each proposed component, and an efficiency comparison. The difficulty-guided attention idea is interesting. However, the central quantitative evidence is undermined by internally impossible metric pairs in the main comparison tables and by a comparison protocol that may handicap the baselines. The claimed state-of-the-art status is therefore not currently supported by the reported evidence, even though the method itself may be salvageable after a careful re-evaluation.

major comments (4)
  1. [Tables 1 and 2] Several reported DSC/mIoU pairs violate the definitional relationship between the two metrics. For any binary prediction, DSC = 2*IoU/(1+IoU), so IoU = DSC/(2-DSC) for each image and the same relationship holds for any consistent averaging over images. In Table 1, MVSSNet on MISD reports DSC 53.9 and mIoU 17.4; a DSC of 53.9 implies an IoU of about 36.9, and an mIoU of 17.4 would imply a DSC of about 29.6. In Table 2, MT-SENet on IMD2020 reports DSC 1.9 and mIoU 12.4, but an mIoU of 12.4 implies a DSC of about 22.1, which is impossible alongside DSC 1.9. Also, the MT-SENet rows for CoMoFoD and "In the Wild" in Table 1 are numerically identical (DSC 10.6, mIoU 6.5), suggesting a copy or paste error. These entries are not rounding artifacts. Because Tables 1 and 2 are the sole quantitative support for the headline outperformance claim, all entries must be recomputed and verified before the claim can be assessed.
  2. [§3.3, Eq. (6)] The curvature formula is incorrect as written. The standard implicit-curvature expression is (G_x^2 G_yy - 2 G_x G_y G_xy + G_y^2 G_xx) / (G_x^2 + G_y^2)^{1.5}. Equation (6) omits the G_xy cross term entirely and writes the middle term as -2 G_x G_y. Consequently the numerator is dimensionally inconsistent: the first and third terms scale as intensity^3/length^4, while the middle term scales as intensity^2/length^2, and the denominator scales as intensity^3/length^3. Since this curvature value is the sole input to the "hard/easy" decision in Algorithm 1, the difficulty signal used by the DGA module is not the quantity the paper claims it to be. The formula and Algorithm 1 should be corrected, and the experiments re-run or at least checked for sensitivity to this implementation detail.
  3. [§4.1] The fairness of the baseline comparison is not established. The paper states that all models were trained with the same batch size (32), learning rate (1e-4), optimizer (Adam), schedule (cosine annealing), resolution (256x256), and 100-epoch budget. Many of the compared methods, including MVSSNet, EITLNet, and PIMNet, were published with specialized multi-task losses, multi-scale supervision, or higher-resolution inputs, and a single generic schedule can put them at a systematic disadvantage. Identical settings are not the same as fair settings. The authors should report the per-method hyperparameters actually used, state whether each baseline's published training recipe was followed or whether a tuning budget was applied, and provide the validation-based model selection criterion. Without this, the reported margin over baselines cannot be attributed to the M2SFormer design.
  4. [§3.3 and Table 4] The difficulty label is endogenously derived from the model's own global prior map G (Algorithm 1), so early in training the "hard"/"easy" signal reflects the model's current errors rather than an independent property of the sample. The paper provides no control experiment, such as using the ground-truth mask to compute difficulty or replacing the label with a random binary signal, to show that the DGA mechanism itself is what improves performance. Moreover, Table 4 shows that "Simple DC + DGA" decreases unseen-domain DSC relative to "No DGA" (30.8 vs. 32.3), while "EADC + DGA" improves it (43.0). This non-monotonic result requires an explanation; otherwise the large EADC gain may come from the edge-aware computation rather than from the difficulty-guided attention, or from an interaction that the current experiments do not isolate.
minor comments (7)
  1. [Contributions, §1] There is a typo "that that efficiently integrates" in the first contribution bullet.
  2. [Table 1] The MT-SENet row appears to be duplicated for CoMoFoD and In the Wild; this should be corrected even if the underlying numbers are recomputed.
  3. [§3.2, Eq. (2)] The DCT basis image definition uses H_s and W_s, but these are not defined; they should presumably be H_t and W_t, and the top-K selection of (u_k, v_k) should be described more concretely.
  4. [§3.2, Eq. (3)] The notation is inconsistent: f^k_c is defined as a per-channel DCT coefficient vector, but Eq. (3) applies spatial pooling and 1x1 convolutions to it as though it were a feature map. Please clarify the tensor shapes and the exact operations.
  5. [Appendix §10, Eq. (9)] Equation (9) defines per-image IoU, not mean IoU; the averaging over images should be stated explicitly, and the same applies to DSC in Eq. (8).
  6. [Table 4 caption] The caption spells the acronym as "ECDC" while the text and setting name use "EADC"; please make the notation consistent.
  7. [Figure 4 caption] There is a typo "qualtitative" in the figure caption; also, the supplementary dataset description uses "MSID" instead of "MISD" in one place.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; minor self-citation to prior M2S work is not load-bearing.

full rationale

The paper's claimed derivation chain is empirical and self-contained: M2SFormer is trained end-to-end on CASIAv2 or DIS25k with a combined BCE loss over the final prediction and the global prior map (Section 3.3), then evaluated on six held-out datasets (Section 4.1). No equation reduces a reported metric to a fitted parameter or to the method's own definition. The only self-citation of note is [50] (same first author), cited alongside [18,56,60] as prior multi-spectral attention work; the paper explicitly differentiates its 'convolution-based statistical aggregation block' and supports the component with an ablation (Table 3), so the citation is positioning rather than load-bearing. The DGA difficulty signal is a heuristic computed from the model's own predicted prior map (Eq. 6, Algorithm 1) and is trained with a separate BCE term; it is not a definition of the target result. I therefore find no circular step. Separately, the reported comparison tables contain arithmetically impossible DSC/mIoU pairs (Table 1 MVSSNet/MISD: DSC 53.9, mIoU 17.4; Table 2 MT-SENet/IMD2020: DSC 1.9, mIoU 12.4) that violate the standard DSC/IoU relation; this is a correctness risk in the empirical evidence, outside the circularity question.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The architecture's claimed gains rest on a small set of hand-set hyperparameters (tau, K/F, L, C_r, C_T, target resolution) and on domain assumptions about curvature-based difficulty, ImageNet transfer, and evaluation protocol fairness. No code or model is released to independently confirm the sensitivity to these choices.

free parameters (6)
  • Difficulty threshold tau = 0.5
    Hand-set threshold in the Edge-Aware Difficulty Calculator that splits samples into 'hard' vs 'easy'; no sensitivity analysis is provided and it directly controls the DGA text token.
  • Number of DCT frequency components (K/F) = F=16 (K not explicitly defined)
    Top-K selection for the 2D DCT basis in Eq. (2); K is not defined explicitly, and F=16 is listed in hyperparameters. Affects the multi-spectral attention.
  • Number of pyramid levels L = 3
    Number of downsampled levels in the multi-scale attention block (Section 3.2, Eq. (4)).
  • Uniform channel C_r = 64
    Channel reduction dimension for skip-connection features (Section 3.2, Eq. (1)); controls model size and is chosen for efficiency.
  • Text embedding dimension C_T = 300
    Dimension of the 'hard'/'easy' text embedding in DGA (Section 3.3, Eq. (7)).
  • Target resolution scale = (H/8, W/8)
    Resolution to which skip-connection features are resized (Section 3.2); with H,W=256, this is 32x32. A hand-chosen hyperparameter.
assumptions (6)
  • standard math 2D DCT basis completeness with top-K selection preserves discriminative frequency components.
    Eq. (2) applies 2D DCT to the cross-scale feature map; top-K selection is inherited from FCANet [56] without re-validation in this setting.
  • domain assumption Curvature of the predicted global prior map correlates with image-level forgery localization difficulty.
    Section 3.3 cites cognitive studies [3,34,74] to argue high-curvature regions demand more attention, then assumes the curvature of the model's own predicted mask is a valid difficulty signal; this link is not empirically measured.
  • domain assumption ImageNet-1K pretrained PVT-v2 features transfer to pixel-level forgery localization.
    Section 3.1 selects PVT-v2 pretrained on ImageNet-1K as backbone; the transferability assumption is standard but not verified for forensics artifacts.
  • domain assumption Resizing all images to 256x256 preserves subtle forgery traces for all methods.
    Section 4.1 mandates 256x256 inputs for every model; high-frequency tampering clues may be destroyed at this resolution, potentially affecting frequency-based baselines.
  • domain assumption The six external test datasets are disjoint from training and representative of the target domain.
    Section 4.1 states these datasets are not used for training; the paper provides no analysis of label noise or protocol leakage.
  • ad hoc to paper A one-bit difficulty label ('hard' vs 'easy') captures enough per-sample difficulty to boost decoder attention.
    The DGA module (Section 3.3) compresses all difficulty variation into a single thresholded token, an ad hoc design choice with no theoretical grounding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M2SFormer: Multi-Spectral and Multi-Scale Attention with Edge-Aware Difficulty Guidance for Image Forgery Localization." pith.science (2026). https://pith.science/paper/TG5RDOSD

@misc{pith2026250620922,
  author       = {Pith},
  title        = {Pith review of: M2SFormer: Multi-Spectral and Multi-Scale Attention with Edge-Aware Difficulty Guidance for Image Forgery Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TG5RDOSD}},
  note         = {Machine review of arXiv:2506.20922}
}
read the original abstract

Image editing techniques have rapidly advanced, facilitating both innovative use cases and malicious manipulation of digital images. Deep learning-based methods have recently achieved high accuracy in pixel-level forgery localization, yet they frequently struggle with computational overhead and limited representation power, particularly for subtle or complex tampering. In this paper, we propose M2SFormer, a novel Transformer encoder-based framework designed to overcome these challenges. Unlike approaches that process spatial and frequency cues separately, M2SFormer unifies multi-frequency and multi-scale attentions in the skip connection, harnessing global context to better capture diverse forgery artifacts. Additionally, our framework addresses the loss of fine detail during upsampling by utilizing a global prior map, a curvature metric indicating the difficulty of forgery localization, which then guides a difficulty-guided attention module to preserve subtle manipulations more effectively. Extensive experiments on multiple benchmark datasets demonstrate that M2SFormer outperforms existing state-of-the-art models, offering superior generalization in detecting and localizing forgeries across unseen domains.

Figures

Figures reproduced from arXiv: 2506.20922 by the authors.

Figure 1
Figure 1. (a) The overall architecture of the proposed [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The overall block diagram of the Difficulty-guided At [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of parameters (M), and FLOPs (G) vs seg [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of other methods and M2SFormer with CASIAv2 training scheme. (a) Input images with ground truth. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of other methods and M2SFormer with DIS25k training scheme. (a) Input images with ground truth. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

84 extracted references · 69 canonical work pages

  1. [1]

    Discrete cosine transform.IEEE transactions on Computers, 100(1): 90–93, 1974

    Nasir Ahmed, T Natarajan, and Kamisetty R Rao. Discrete cosine transform.IEEE transactions on Computers, 100(1): 90–93, 1974. 2

    1974

  2. [2]

    Perception and de- ception: Exploring individual responses to deepfakes across different modalities.Heliyon, 9(10), 2023

    Saifuddin Ahmed and Hui Wen Chua. Perception and de- ception: Exploring individual responses to deepfakes across different modalities.Heliyon, 9(10), 2023. 1

    2023

  3. [3]

    Some informational aspects of visual percep- tion.Psychological review, 61(3):183, 1954

    Fred Attneave. Some informational aspects of visual percep- tion.Psychological review, 61(3):183, 1954. 4

    1954

  4. [4]

    Deep frequency re- calibration u-net for medical image segmentation

    Reza Azad, Afshin Bozorgpour, Maryam Asadi-Aghbolaghi, Dorit Merhof, and Sergio Escalera. Deep frequency re- calibration u-net for medical image segmentation. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 3274–3283, 2021. 1

    2021

  5. [5]

    Segnet: A deep convolutional encoder-decoder architecture for image segmentation.IEEE transactions on pattern anal- ysis and machine intelligence, 39(12):2481–2495, 2017

    Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation.IEEE transactions on pattern anal- ysis and machine intelligence, 39(12):2481–2495, 2017. 6, 7, 1, 3

    work page 2017

  6. [6]

    Pim-net: Progressive in- consistency mining network for image manipulation local- ization.Pattern Recognition, 159:111136, 2025

    Ningning Bai, Xiaofeng Wang, Ruidong Han, Jianpeng Hou, Yihang Wang, and Shanmin Pang. Pim-net: Progressive in- consistency mining network for image manipulation local- ization.Pattern Recognition, 159:111136, 2025. 2, 6, 7, 1, 3

    work page 2025

  7. [7]

    Rru-net: The ringed residual u-net for image splicing forgery detec- tion

    Xiuli Bi, Yang Wei, Bin Xiao, and Weisheng Li. Rru-net: The ringed residual u-net for image splicing forgery detec- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops, pages 0–0,

    work page

  8. [8]

    Image manipulation detection by multi-view multi-scale supervision

    Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image manipulation detection by multi-view multi-scale supervision. InProceedings of the IEEE/CVF international conference on computer vision, pages 14185–14193, 2021. 1

    work page 2021

Show all 84 references
  1. [9]

    Drop an octave: Reducing spatial redundancy in con- volutional neural networks with octave convolution

    Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yan- nis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Jiashi Feng. Drop an octave: Reducing spatial redundancy in con- volutional neural networks with octave convolution. InPro- ceedings of the IEEE/CVF international conferenc...

    2019

  2. [10]

    A bayesian-mrf approach for prnu-based image forgery detection.IEEE Transactions on Information Forensics and Security, 9(4):554–567, 2014

    Giovanni Chierchia, Giovanni Poggi, Carlo Sansone, and Luisa Verdoliva. A bayesian-mrf approach for prnu-based image forgery detection.IEEE Transactions on Information Forensics and Security, 9(4):554–567, 2014. 1, 2

    2014

  3. [11]

    Attentional feature fusion

    Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. InProceed- ings of the IEEE/CVF winter conference on applications of computer vision, pages 3560–3569, 2021. 1

    2021

  4. [12]

    Image tamper detec- tion based on demosaicing artifacts

    Ahmet Emir Dirik and Nasir Memon. Image tamper detec- tion based on demosaicing artifacts. In2009 16th IEEE In- ternational Conference on Image Processing (ICIP), pages 1497–1500. IEEE, 2009. 1, 2

    2009

  5. [13]

    Mvss-net: Multi-view multi-scale supervised net- works for image manipulation detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3539– 3553, 2022

    Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised net- works for image manipulation detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3539– 3553, 2022. 6, 7, 1, 3

    2022

  6. [14]

    CASIA image tam- pering detection evaluation database

    Jing Dong, Wei Wang, and Tieniu Tan. CASIA image tam- pering detection evaluation database. In2013 IEEE China Summit and International Conference on Signal and Infor- mation Processing. IEEE, 2013. 6, 1

    2013

  7. [15]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  8. [16]

    Image forgery localization via fine-grained analysis of cfa artifacts.IEEE Transactions on Information Forensics and Security, 7(5):1566–1577, 2012

    Pasquale Ferrara, Tiziano Bianchi, Alessia De Rosa, and Alessandro Piva. Image forgery localization via fine-grained analysis of cfa artifacts.IEEE Transactions on Information Forensics and Security, 7(5):1566–1577, 2012. 1, 2

    2012

  9. [17]

    Res2net: A new multi-scale backbone architecture.IEEE transactions on pattern analysis and machine intelligence, 43(2):652–662,

    Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture.IEEE transactions on pattern analysis and machine intelligence, 43(2):652–662,

  10. [18]

    Fbi-net: Frequency-based image forgery localization via multitask learning with self-attention.IEEE Access, 10:62751–62762,

    A-Rom Gu, Ju-Hyeon Nam, and Sang-Chul Lee. Fbi-net: Frequency-based image forgery localization via multitask learning with self-attention.IEEE Access, 10:62751–62762,

  11. [19]

    Effective im- age tampering localization via enhanced transformer and co-attention fusion

    Kun Guo, Haochen Zhu, and Gang Cao. Effective im- age tampering localization via enhanced transformer and co-attention fusion. InICASSP 2024-2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 4895–4899. IEEE, 2024. 2, 6, 7, 1, 3

    2024

  12. [20]

    Segnext: Rethink- ing convolutional attention design for semantic segmenta- tion.Advances in Neural Information Processing Systems, 35:1140–1156, 2022

    Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethink- ing convolutional attention design for semantic segmenta- tion.Advances in Neural Information Processing Systems, 35:1140–1156, 2022. 6, 7, 1, 2, 3

    2022

  13. [21]

    Hierarchical fine-grained im- age forgery detection and localization

    Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Ia- copo Masi, and Xiaoming Liu. Hierarchical fine-grained im- age forgery detection and localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155–3165, 2023. 1

    2023

  14. [22]

    Hdf-net: Capturing ho- mogeny difference features to localize the tampered image

    Ruidong Han, Xiaofeng Wang, Ningning Bai, Yihang Wang, Jianpeng Hou, and Jianru Xue. Hdf-net: Capturing ho- mogeny difference features to localize the tampered image. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2024. 2

    2024

  15. [23]

    Transforensics: image forgery localization with dense 9 self-attention

    Jing Hao, Zhixin Zhang, Shicai Yang, Di Xie, and Shiliang Pu. Transforensics: image forgery localization with dense 9 self-attention. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15055–15064, 2021. 1, 2, 6, 7, 3

    2021

  16. [24]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2

    2016

  17. [25]

    Bpemb: Tokenization-free pre-trained subword embeddings in 275 languages.arXiv preprint arXiv:1710.02187, 2017

    Benjamin Heinzerling and Michael Strube. Bpemb: Tokenization-free pre-trained subword embeddings in 275 languages.arXiv preprint arXiv:1710.02187, 2017. 5

    2017 arXiv

  18. [26]

    Hsu and S.-F

    Y .-F. Hsu and S.-F. Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. InInternational Conference on Multimedia and Expo, 2006. 6, 1

    2006

  19. [27]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018. 1, 2

    2018

  20. [28]

    Span: Spa- tial pyramid attention network for image manipulation local- ization

    Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spa- tial pyramid attention network for image manipulation local- ization. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceed- ings, ...

    2020

  21. [29]

    Fighting fake news: Image splice detection via learned self-consistency

    Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. InProceedings of the European conference on computer vision (ECCV), pages 101–117,

  22. [30]

    Multiple image splicing dataset (misd): a dataset for multiple splicing.Data, 6(10):102, 2021

    Kalyani Dhananjay Kadam, Swati Ahirrao, and Ketan Kotecha. Multiple image splicing dataset (misd): a dataset for multiple splicing.Data, 6(10):102, 2021. 6, 1

    2021

  23. [31]

    Design of an image edge detection filter using the sobel operator.IEEE Journal of solid-state circuits, 23(2): 358–367, 1988

    Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator.IEEE Journal of solid-state circuits, 23(2): 358–367, 1988. 5

    1988

  24. [32]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007–6017, 2023. 1

    2023

  25. [33]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    arXiv

  26. [34]

    The structure of images.Biological cy- bernetics, 50(5):363–370, 1984

    Jan J Koenderink. The structure of images.Biological cy- bernetics, 50(5):363–370, 1984. 4

    1984

  27. [35]

    Multi-scale analysis strategies in prnu-based tampering localization.IEEE Transactions on Information Forensics and Security, 12(4):809–824, 2016

    Paweł Korus and Jiwu Huang. Multi-scale analysis strategies in prnu-based tampering localization.IEEE Transactions on Information Forensics and Security, 12(4):809–824, 2016. 1, 2

    2016

  28. [36]

    Diffusion- based conditional image editing through optimized inference with guidance.arXiv preprint arXiv:2412.15798, 2024

    Hyunsoo Lee, Minsoo Kang, and Bohyung Han. Diffusion- based conditional image editing through optimized inference with guidance.arXiv preprint arXiv:2412.15798, 2024. 1

    2024 arXiv

  29. [37]

    Wavesnet: Wavelet integrated deep networks for image segmentation

    Qiufu Li and Linlin Shen. Wavesnet: Wavelet integrated deep networks for image segmentation. InChinese Confer- ence on Pattern Recognition and Computer Vision (PRCV), pages 325–337. Springer, 2022. 1

    2022

  30. [38]

    Wavelet integrated cnns for noise-robust image classification

    Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavelet integrated cnns for noise-robust image classification. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7245–7254, 2020. 1

    2020

  31. [39]

    Frequency-aware camouflaged object detection.ACM Transactions on Multimedia Computing, Communications and Applications, 19(2):1–16, 2023

    Jiaying Lin, Xin Tan, Ke Xu, Lizhuang Ma, and Rynson WH Lau. Frequency-aware camouflaged object detection.ACM Transactions on Multimedia Computing, Communications and Applications, 19(2):1–16, 2023. 1

    2023

  32. [40]

    Referring image editing: Object-level image editing via referring expressions

    Chang Liu, Xiangtai Li, and Henghui Ding. Referring image editing: Object-level image editing via referring expressions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13128–13138, 2024. 1

    2024

  33. [41]

    Attention consistency refined masked frequency forgery representation for generalizing face forgery detection.IEEE Transactions on Information Forensics and Security, 2024

    Decheng Liu, Tao Chen, Chunlei Peng, Nannan Wang, Ruimin Hu, and Xinbo Gao. Attention consistency refined masked frequency forgery representation for generalizing face forgery detection.IEEE Transactions on Information Forensics and Security, 2024. 1

    2024

  34. [42]

    Spatial- phase shallow learning: rethinking face forgery detection in frequency domain

    Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial- phase shallow learning: rethinking face forgery detection in frequency domain. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pag...

    2021

  35. [43]

    Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pscc-net: Progressive spatio-channel correlation network for image manipulation detection and localization.IEEE Trans- actions on Circuits and Systems for Video Technology, 32 (11):7505–7517, 2022. 1, 2

    2022

  36. [44]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 3

    2021

  37. [45]

    Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016. 6

    2016 arXiv

  38. [46]

    Distinctive image features from scale- invariant keypoints.International journal of computer vi- sion, 60:91–110, 2004

    David G Lowe. Distinctive image features from scale- invariant keypoints.International journal of computer vi- sion, 60:91–110, 2004. 2, 3, 4

    2004

  39. [47]

    Gener- alizing face forgery detection with high-frequency features

    Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Gener- alizing face forgery detection with high-frequency features. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 16317–16326, 2021. 1

    2021

  40. [48]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In2016 fourth international conference on 3D vision (3DV), pages 565–571. Ieee, 2016. 6, 2

    2016

  41. [49]

    A survey on the detection and impacts of deepfakes in visual, audio, and tex- tual formats.Ieee Access, 11:144497–144529, 2023

    Rami Mubarak, Tariq Alsboui, Omar Alshaikh, Isa Inuwa- Dutse, Saad Khan, and Simon Parkinson. A survey on the detection and impacts of deepfakes in visual, audio, and tex- tual formats.Ieee Access, 11:144497–144529, 2023. 1

    2023

  42. [50]

    Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multi- scale attention

    Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multi- scale attention. InProceedings of the IEEE/CVF Conference 10 on Computer Vision and Pattern Recognition, pages 1148...

    2024

  43. [51]

    Cfl-net: image forgery localization using contrastive learning

    Fahim Faisal Niloy, Kishor Kumar Bhaumik, and Simon S Woo. Cfl-net: image forgery localization using contrastive learning. InProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 4642–4651,

  44. [52]

    Imd2020: A large-scale annotated dataset tailored for de- tecting manipulated images

    Adam Novozamsky, Babak Mahdian, and Stanislav Saic. Imd2020: A large-scale annotated dataset tailored for de- tecting manipulated images. In2020 IEEE Winter Applica- tions of Computer Vision Workshops (WACVW), pages 71– 80, 2020. 6, 1

    2020

  45. [53]

    Hybrid image-retrieval method for image- splicing validation.Symmetry, 11(1):83, 2019

    Nam Thanh Pham, Jong-Weon Lee, Goo-Rak Kwon, and Chun-Su Park. Hybrid image-retrieval method for image- splicing validation.Symmetry, 11(1):83, 2019. 6, 8, 1, 2

    2019

  46. [54]

    Statistical tools for digital forensics

    Alin C Popescu and Hany Farid. Statistical tools for digital forensics. InInternational workshop on information hiding, pages 128–147. Springer, 2004. 1, 2

    2004

  47. [55]

    Thinking in frequency: Face forgery detection by min- ing frequency-aware clues

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by min- ing frequency-aware clues. InEuropean conference on com- puter vision, pages 86–103. Springer, 2020. 1

    2020

  48. [56]

    Fcanet: Frequency channel attention networks

    Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. InProceedings of the IEEE/CVF international conference on computer vision, pages 783–792, 2021. 1, 4

    2021

  49. [57]

    Digital image forensics: a booklet for beginners.Multimedia Tools and Applications, 51:133–162, 2011

    Judith A Redi, Wiem Taktak, and Jean-Luc Dugelay. Digital image forensics: a booklet for beginners.Multimedia Tools and Applications, 51:133–162, 2011. 3

    2011

  50. [58]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, par...

    2015

  51. [59]

    Imagenet large scale visual recognition challenge.International journal of computer vision, 115:211–252, 2015

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge.International journal of computer vision, 115:211–252, 2015. 3

    2015

  52. [60]

    Multi-frequency informa- tion enhanced channel attention module for speaker repre- sentation learning.arXiv preprint arXiv:2207.04540, 2022

    Mufan Sang and John HL Hansen. Multi-frequency informa- tion enhanced channel attention module for speaker repre- sentation learning.arXiv preprint arXiv:2207.04540, 2022. 4

    2022 arXiv

  53. [61]

    Deep image composition meets image forgery, 2024

    Eren Tahir and Mert Bal. Deep image composition meets image forgery, 2024. 6, 1

    2024

  54. [62]

    Comofod—new database for copy-move forgery detection

    Dijana Tralic, Ivan Zupancic, Sonja Grgic, and Mislav Grgic. Comofod—new database for copy-move forgery detection. InProceedings ELMAR-2013, pages 49–54. IEEE, 2013. 6, 1

    2013

  55. [63]

    Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news.Social media+ society, 6(1):2056305120903408, 2020

    Cristian Vaccari and Andrew Chadwick. Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news.Social media+ society, 6(1):2056305120903408, 2020. 1

    2020

  56. [64]

    Ob- jectformer for image manipulation detection and localiza- tion

    Junke Wang, Zuxuan Wu, Jingjing Chen, Xintong Han, Ab- hinav Shrivastava, Ser-Nam Lim, and Yu-Gang Jiang. Ob- jectformer for image manipulation detection and localiza- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 2364–2373,

  57. [65]

    Pvt v2: Improved baselines with pyramid vision transformer

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415–424, 2022. 3, 2

    2022

  58. [66]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 1

    2018

  59. [67]

    Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features

    Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 9543–9552, 2019....

    2019

  60. [68]

    P2t: Pyramid pooling transformer for scene understanding

    Yu-Huan Wu, Yun Liu, Xin Zhan, and Ming-Ming Cheng. P2t: Pyramid pooling transformer for scene understanding. IEEE transactions on pattern analysis and machine intelli- gence, 2022. 2

    2022

  61. [69]

    Image tampering detection with frequency-aware at- tention and multi-view fusion.IEEE Transactions on Artifi- cial Intelligence, 2024

    Xu Xu, Junxin Chen, Wenrui Lv, Wei Wang, and Yushu Zhang. Image tampering detection with frequency-aware at- tention and multi-view fusion.IEEE Transactions on Artifi- cial Intelligence, 2024. 1, 2

    2024

  62. [70]

    Paint by example: Exemplar-based image editing with diffusion mod- els

    Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion mod- els. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 18381–18391,

  63. [71]

    Jianquan Yang, Yulan Zhang, Guopu Zhu, and Sam Kwong. A clustering-based framework for improving the perfor- mance of jpeg quantization step estimation.IEEE trans- actions on circuits and systems for video technology, 31(4): 1661–1672, 2020. 1, 2

    2020

  64. [72]

    Fda: Fourier domain adaptation for semantic segmentation

    Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085–4095, 2020. 1

    2020

  65. [73]

    D-net: A dual-encoder net- work for image splicing forgery detection and localization

    Zonglin Yang, Bo Liu, Xiuli Bi, Bin Xiao, Weisheng Li, Guoyin Wang, and Xinbo Gao. D-net: A dual-encoder net- work for image splicing forgery detection and localization. Pattern Recognition, 155:110727, 2024. 2

    2024

  66. [74]

    Cur- vature processing in human visual cortical areas.NeuroIm- age, 222:117295, 2020

    Xiaomin Yue, Sophia Robert, and Leslie G Ungerleider. Cur- vature processing in human visual cortical areas.NeuroIm- age, 222:117295, 2020. 4

    2020

  67. [75]

    Wcanet: Wavelet channel attention network for citrus variety identification

    Fukai Zhang, Xiaobo Jin, Jie Jiang, Shan An, and Qiang Lyu. Wcanet: Wavelet channel attention network for citrus variety identification. In2023 IEEE International Conference on Image Processing (ICIP), pages 2845–2849. IEEE, 2023. 1

    2023

  68. [76]

    Resnest: Split-attention networks

    Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 2736–2746, 2022. 2

    2022

  69. [77]

    Multi-task se-network for image 11 splicing localization.IEEE Transactions on Circuits and Sys- tems for Video Technology, 32(7):4828–4840, 2021

    Yulan Zhang, Guopu Zhu, Ligang Wu, Sam Kwong, Hongli Zhang, and Yicong Zhou. Multi-task se-network for image 11 splicing localization.IEEE Transactions on Circuits and Sys- tems for Video Technology, 32(7):4828–4840, 2021. 1, 2, 6, 7, 3

    2021

  70. [78]

    Detecting camouflaged object in fre- quency domain

    Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in fre- quency domain. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4504– 4513, 2022. 1

    2022

  71. [79]

    Learning rich features for image manipulation detection

    Peng Zhou, Xintong Han, Vlad I Morariu, and Larry S Davis. Learning rich features for image manipulation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1053–1061, 2018. 2 12 M2SFormer: Multi-Spectral and Multi-Scale Attention...

    2018

  72. [80]

    CASIAv2 is more complex than CASIAv1, containing 5,123 tampered ones which consists of 3274 copy move images and 1849 splicing images

    Dataset Descriptions •CASIAv1[14] andCASIAv2[53]: The CASIAv1 dataset consists of JPG images with a resolution of 384 × 256, including 459 copy move and 461 splicing images. CASIAv2 is more complex than CASIAv1, containing 5,123 tampered ones which consists of 3274 copy move i...

    1920

  73. [81]

    Technical Novelty of M2SFormer M2SFormerintroduces a unified framework that integrates multi-frequency and multi-scale attention in a single stream, addressing a longstanding challenge in forgery localiza- tion where frequency- and spatial-domain features were traditionally pr...

  74. [82]

    Broader Impact in Artificial Intelligence M2SFormer’s unified approach to forgery localiza- tion—combiningmulti-spectral and multi-scale attention with thetext-guided difficulty attention—not only enhances detection accuracy for unseen or subtle manipulations but also carries ...

  75. [83]

    More Detailed Ablation Study on M2SFormer In this section, we perform a more detailed ablation study on M2SFormer. 9.1. Ablation Study on Backbone in M2SFormer In this section, we conduct an ablation study to evaluate the impact of different backbone models on the performance ...

  76. [84]

    SP” and “CM

    Metrics Descriptions In this section, we describe the metrics used in this paper. For convenience, we denoteT P, F P, andF Nas the num- ber of samples of true positive, false positive, and false neg- ative between two binary masksAandB. • TheMean Dice Similarity Coefficient (D...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.