Pith. sign in

REVIEW 4 major objections 5 minor 38 references

EfficientIML: Efficient High-Resolution Image Manipulation Localization

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A linear-cost RWKV backbone and a high-resolution semantic-inpainting benchmark push image-manipulation localization to 2048×2048 inputs, where the new model reports 78.0 pixel F1, 13 points above the strongest compared baseline.

desk verdict Useful dataset and a genuinely efficient backbone, but the SIF benchmark's auto-segmented masks let a model win by doing objectness, so the headline +13 F1 needs a control baseline before it means what they claim. read the letter →

arxiv 2509.08583 v1 pith:O5EU4KEW submitted 2025-09-10 cs.CV

Jinhan Li , Haoyang He , Lei Xie , Jiangning Zhang This is my paper
classification cs.CV
keywords imagemanipulationlocalizationdiffusionforgerydetectionsemanticinpaintinghigh-resolutiondatasetRWKVvisionbackbonelightweightforensicsmulti-scalesupervision
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 argues that image-manipulation localization can be pushed to today's high resolutions without the computational blowup of vision transformers, and that diffusion-generated inpainting forgeries need a dedicated high-resolution benchmark. To make that case, it introduces SIF, a dataset of 1,228 diffusion-inpainted images at up to 1800×1200 pixels with semantically derived ground-truth masks. It then proposes EfficientIML, built on a three-stage EfficientRWKV backbone whose global state-space branch, local depthwise-convolution branch, and identity branch share the work at linear cost. On the SIF test split at 2048×2048, the model reaches 78.0 F1, 13 points above the strongest baseline, while using roughly a quarter of the lightest baseline's FLOPs. If the dataset is representative, this suggests real-time forensic screening of high-resolution images is within reach.

What carries the argument

The load-bearing object is the EfficientRWKV block and the three-stage backbone built from it. It routes a small channel slice through Vision-RWKV6's Bi-WKV recurrence, a linear-complexity global attention that decays past tokens rather than attending pairwise; routes another slice through a depthwise convolutional block sized by stage; and leaves the remaining channels as an identity shortcut. This division is what lets the model capture global context and local detail without quadratic cost. A multi-scale decoder with per-scale BCE and edge losses turns the three backbone feature maps into the final mask, and the ablations tie the SIF gains to the global-plus-local branch combination and t

What would settle it

Run the same model comparison on a high-resolution set of diffusion forgeries with human-drawn ground-truth masks, or on all 5,000 uncurated generated images; if EfficientIML's 13-point F1 lead over MVSS shrinks or reverses, the reported SIF advantage is an artifact of the automatic masks or curation.

Watch

Extended reading notes

Core claim

The central claim is that a deliberately lightweight recurrent-style backbone can localize manipulated regions in high-resolution images more accurately than heavier transformer and CNN detectors. EfficientRWKV replaces pairwise attention with a three-branch block: a fixed channel budget runs the linear WKV recurrence for global context, a second channel group runs depthwise convolutions for fine local patterns, and the remaining channels pass through unchanged. With this block arranged in three downsampling stages and supervised at each scale, the model reports the best F1, AUC, IoU, and accuracy on the new SIF benchmark and on standard cross-dataset tests, while its complexity stays under

Load-bearing premise

The SIF benchmark is a fair test of real diffusion forgeries: its masks come from an automatic segmenter, and its 1,228 images were hand-picked from 5,000 generated candidates by subjective quality criteria.

Editorial extensions

If this is right

  • At 1024×1024, EfficientIML uses about 21.7 GFLOPs versus 95.6 for SHViT, the lightest compared model, while reporting higher F1 on SIF (68.1 vs 55.9).
  • At 2048×2048 the model keeps 86.7 GFLOPs and 23.0 images per second, versus 382.4 GFLOPs and 16.9 for SHViT and 2617.8 GFLOPs and 2.4 for MVSS.
  • On the CASIA V2 cross-dataset protocol, the method raises average pixel F1 to 42.6 from 38.3 for the best prior baseline, with the largest margins on CASIA V1 and Columbia.
  • SIF gives the field a common testbed above 1024×1024 for diffusion-based semantic inpainting, complementing lower-resolution diffusion datasets such as CSI-IMD.

Reading between the lines

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

  • Because SIF's masks come from HIPPIE's segmentation outputs, a sharper test would retrain on SIF with human-drawn masks or with masks produced by several segmenters; if the 13-point margin shrinks, part of the gain is learning the segmenter's contours rather than the forgery itself.
  • The linear-complexity backbone suggests a direct extension to 4K and possibly video frames, where quadratic ViT attention becomes prohibitive; the paper does not run those experiments.
  • A lightweight, real-time localizer could be embedded in cameras or upload pipelines as a first-pass forensic filter; the paper mentions real-time suitability but does not test on-device deployment.
  • Since the model is trained only on Stable Diffusion 2 inpaints, its performance on other generative editors is untested; an obvious extension is a mixed-source training set covering newer diffusion models and GAN inpainting.
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 / 5 minor

Summary. The paper introduces SIF, a new dataset of 1,228 high-resolution diffusion-based inpainting forgeries, and EfficientIML, a lightweight three-stage backbone based on EfficientRWKV with a multi-scale decoder. The dataset is constructed by using HIPPIE to obtain semantic object masks, Qwen-VL to generate inpainting prompts, and Stable Diffusion to replace the masked objects. The method is evaluated under two protocols: training on CASIA V2 and testing on six standard benchmarks, and training/testing on SIF at 1024x1024 and 2048x2048. The paper reports consistent gains over lightweight baselines in F1/AUC/IoU/ACC and substantially lower FLOPs and higher inference speed, arguing suitability for real-time high-resolution forensics.

Significance. If the claims hold, the paper makes two useful contributions: a high-resolution semantic inpainting forgery dataset and an efficient linear-complexity backbone that outperforms existing lightweight detectors on standard benchmarks. The CASIA V2-trained cross-dataset evaluation (Table 1) and the efficiency measurements (Table 3) are concrete, reproducible strengths. However, the SIF-specific performance claims, especially the headline +13 F1 gain over MVSS at 2048x2048 (Table 2b), rest on a dataset whose ground-truth masks are generated by automatic semantic segmentation rather than human annotation, creating a potential objectness shortcut. This is a load-bearing concern that must be addressed before the SIF-based conclusions can be accepted.

major comments (4)
  1. [Section 2, Table 2b, Appendix A] The SIF ground-truth masks are obtained by HIPPIE, an automatic segmentation model, and the manipulated region is exactly the semantic object that Stable Diffusion replaces. Appendix A states that 'SIF's context-aware masks trace the contours of actual scene elements' and 'inpainted areas correspond to semantically meaningful objects.' Consequently, a model can achieve high localization scores by performing semantic segmentation or salient-object detection, without detecting any diffusion artifact. The paper reports no control baseline of this kind. The +13 F1 advantage over MVSS in Table 2b (78.0 vs. 65.0 at 2048x2048) may therefore reflect objectness rather than manipulation localization. Please add a baseline such as an off-the-shelf panoptic/semantic segmentation model or salient object detector evaluated on SIF, and/or revise the dataset construction so that masks are not coextensiv
  2. [Section 4.1, Tables 1, 2, 4] All quantitative results are reported from a single training run and a single train/test split. The SIF test set has roughly 175 images (1,228 images split 6:1), so differences of a few F1 points in Table 4 (e.g., 65.7 vs. 65.2 between loss-weight sets, and 68.1 final) may be within noise. The CASIA V2-trained results in Table 1 also lack error bars or significance tests. Please provide results over multiple seeds (at least 3) with mean and standard deviation, or otherwise demonstrate that the reported differences are not due to random variation.
  3. [Section 3.2, Table 8] The chosen loss weights and channel allocations appear to be selected based on the same SIF test set used for the final evaluation in Table 2. This risks overfitting to the test split. The paper should either use a separate validation split for hyperparameter choice or explicitly state the selection procedure. Without this, the SIF-trained performance numbers in Tables 2 and 4 are difficult to interpret as unbiased estimates.
  4. [Section 4.1, Table 3] The efficiency comparison is central to the 'suitable for real-time forensics' claim, but the measurement setup is under-specified: no hardware platform, batch size, or FLOPs-counting method is given, and the use of checkpointing (marked with an asterisk only for ViT-S) may affect speed measurements at 2048x2048. Please report the exact inference setup and confirm that the FLOPs figures include all decoder/auxiliary heads, so that the claimed 86.7 GFLOPs at 2048x2048 is comparable across methods.
minor comments (5)
  1. [Table 1(d)] Column header 'COVERAGE' is typeset inconsistently; use 'Coverage' as in the text and other tables.
  2. [Eq. (1)] The symbol T in the WKV recurrence is used without definition; state that T is the sequence length, and clarify the exact shape of w and u relative to C_v.
  3. [Abstract and Section 2] The abstract says '1200+ diffusion-generated manipulations with semantically extracted masks,' while Section 2 says masks come from HIPPIE. The phrase 'human-validated' in Section 1 refers to image quality, not mask accuracy; this should be clarified to avoid the impression that masks are human-annotated.
  4. [Section 4.2] Reference to 'Table 2b' should be 'Table 2(b)' for consistency with other table references.
  5. [Appendix B.2] Typo: 'Experiemnets' should be 'Experiments'.

Circularity Check

1 steps flagged · score 6.0 of 10

SIF-trained results are partially circular: ground-truth masks equal HIPPIE semantic segmentations, so the SIF localization comparison reduces to semantic segmentation; CASIA-trained results remain independent.

  1. self definitional [Section 2 (SIF dataset construction) and Section 4.2 Table 2b; Appendix Fig. 6]
    "we leverage the Hierarchical Open-vocabulary Universal Image Segmentation[17] (HIPPIE) to segment for the object mask... Then, the prompt, original image, and mask are input into Stable Diffusion 2 [1]. This model replaces the masked region in the image with semantically coherent and visually consistent content. ... SIF's context-aware masks trace the contours of actual scene elements, so the inpainted areas correspond to semantically meaningful objects."

    By construction, the SIF ground-truth mask for each image is the HIPPIE object mask M = HIPPIE(I), and Stable Diffusion inpaints exactly that masked region. Therefore the task 'localize the manipulated region' on SIF is equivalent to 'segment the semantic object.' A model implementing or approximating HIPPIE, or any semantic-segmentation/saliency model, can achieve high F1/AUC/IoU without detecting any diffusion artifact. The Table 2b result (+13% F1 over MVSS, 78.0 vs 65.0) is thus not evidence of manipulation-localization ability; it may simply reflect better semantic segmentation. No semantic-segmentation control baseline is reported, so the SIF-trained performance claim partially reduces to the dataset's self-definitional target.

full rationale

The paper's central efficiency claim and its CASIA V2-trained cross-dataset results (Table 1) are validated against external, human-annotated benchmarks and do not exhibit circularity. The model architecture and loss are stated explicitly and are not derived from the evaluation targets. However, the SIF dataset construction defines the manipulated region as the HIPPIE semantic object mask, and the inpainted area exactly coincides with that mask. Consequently, the SIF-trained evaluation (Tables 2a/2b, especially the +13% F1 headline at 2048x2048) measures agreement with a semantic segmentation output rather than with independently verified manipulation traces. This is a self-definitional circularity in the benchmark: the ground truth is a function of the input via a known segmentation model, so high scores can be achieved by solving a different, easier task. The absence of a semantic-segmentation or saliency baseline makes the SIF-based superiority claim partially circular. The CASIA-trained and efficiency results are independent and should be weighted accordingly.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The model's empirical performance rests on several hand-chosen hyperparameters and on the validity of the auto-generated SIF ground truth. The main external check is the CASIA V2-trained evaluation on standard benchmarks, which partially mitigates the self-referential SIF split.

free parameters (4)
  • Multi-scale loss weights = lambda_1=0.15, lambda_2=0.35, lambda_3=0.55, fused=1.0 (Final Setup, Table 8)
    These weights are chosen by the authors after ablation to optimize SIF validation F1.
  • Edge loss weight lambda_edge = not specified in paper
    Inherited from IML-ViT; not listed.
  • Channel widths per stage = C1=200, C2=376, C3=448 (Appendix B.1)
    Manually set architecture hyperparameters.
  • Branch channel allocation fractions = Stage1: 0.8C/0.2C/0; Stage2: 0.7/0.2/0.1; Stage3: 0.6/0.3/0.1 (Table 7)
    Hand-chosen split of global/local/identity channels.
assumptions (5)
  • domain assumption HIPPIE-segmented object masks are valid ground truth for diffusion inpainting localization
    Section 2: masks are used both to drive inpainting and as ground truth; no manual verification of mask boundaries.
  • domain assumption Human selection of 1,228 images from 5,000 generated yields a representative, unbiased dataset
    Section 2: selection by visual realism, semantic coherence, and diversity; this curation may remove hard cases.
  • domain assumption Stable Diffusion 2 inpainting outputs are representative of real-world diffusion-based forgeries
    Used to create SIF; no external diffusion models tested.
  • domain assumption The IMDL-BenCo implementations of MVSS and PSCC are correct and fairly tuned
    Section 4.1: baselines are run via IMDL-BenCo; training details are not provided.
  • domain assumption RWKV-style linear attention retains accuracy when combined with convolutional branches
    Efficiency and accuracy rely on the Vision-RWKV mechanism; no separate theoretical justification is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EfficientIML: Efficient High-Resolution Image Manipulation Localization." pith.science (2026). https://pith.science/paper/O5EU4KEW

@misc{pith2026250908583,
  author       = {Pith},
  title        = {Pith review of: EfficientIML: Efficient High-Resolution Image Manipulation Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O5EU4KEW}},
  note         = {Machine review of arXiv:2509.08583}
}
read the original abstract

With imaging devices delivering ever-higher resolutions and the emerging diffusion-based forgery methods, current detectors trained only on traditional datasets (with splicing, copy-moving and object removal forgeries) lack exposure to this new manipulation type. To address this, we propose a novel high-resolution SIF dataset of 1200+ diffusion-generated manipulations with semantically extracted masks. However, this also imposes a challenge on existing methods, as they face significant computational resource constraints due to their prohibitive computational complexities. Therefore, we propose a novel EfficientIML model with a lightweight, three-stage EfficientRWKV backbone. EfficientRWKV's hybrid state-space and attention network captures global context and local details in parallel, while a multi-scale supervision strategy enforces consistency across hierarchical predictions. Extensive evaluations on our dataset and standard benchmarks demonstrate that our approach outperforms ViT-based and other SOTA lightweight baselines in localization performance, FLOPs and inference speed, underscoring its suitability for real-time forensic applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 4 linked inside Pith

  1. [1]

    INTRODUCTION Modern forensic analysis increasingly relies on Image Manip- ulation Localization (IML) to detect and precisely mask ma- nipulated regions in digital images. With the development of high-fidelity diffusion-based inpainting methods such as Sta- ble Diffusion[1] and the proliferation of ultra–high–resolution imaging devices, traditional localiz...

    arXiv

  2. [2]

    A lightweight, three–stageEfficientRWKVbackbone that fuses global state–space attention and local convolutional pro- cessing, with linear computational complexities

  3. [3]

    A multi-scale decoder equipped with tailored loss weights

  4. [4]

    Our construction pipeline is a modified version based on the Inpaint Anything framework[15] and the source images are selected from the SAM9K dataset[16]

    HIGH-RESOLUTION SIF DATASET In this study, we introduce a high-resolution manipulated image dataset consisting of 1,228 carefully curated samples. Our construction pipeline is a modified version based on the Inpaint Anything framework[15] and the source images are selected from the SAM9K dataset[16]. The procedure for dataset generation is illustrated in ...

  5. [5]

    3) by integrating an optimized EfficientRWKV backbone with a customized multi-scale prediction scheme

    METHODOLOGY: EFFICIENTIML In this work, we propose a novel and efficient image ma- nipulation localization framework EfficientIML (Fig. 3) by integrating an optimized EfficientRWKV backbone with a customized multi-scale prediction scheme. Our pipeline Fig. 4:EfficientRWKV6 Module builds upon the baseline structure of IML-ViT[11], but in- troduces these tw...

  6. [6]

    EXPERIMENT 4.1. Implementation Details EfficientRWKV .The EfficientRWKV backbone uses a pre- trained checkpoint obtained by pre-training for 1,000 epochs on ImageNet-1K under a teacher–student distillation regime, with a TResNet-L network[20] providing soft targets ac- cording to the DeiT distillation protocol[21]. Training is performed with an effective ...

  7. [7]

    CASIA V2-training: Aligning with IML-ViT[11], we use entire CASIA V2 for training, other datasets for testing

  8. [8]

    Evaluation Criteria.We evaluate all models using four met- rics: Pixel-F1 score, Area Under the ROC Curve (AUC), In- tersection over Union (IoU), and Pixel-Accuracy

    SIF-training: In the second experiment, SIF dataset is split with a 6:1 ratio for training and testing. Evaluation Criteria.We evaluate all models using four met- rics: Pixel-F1 score, Area Under the ROC Curve (AUC), In- tersection over Union (IoU), and Pixel-Accuracy. Table 2:SIF-training at different resolutions (a) Resolution 1024×1024(%) F1 AUC IoU AC...

Show all 38 references
  1. [9]

    Ex- tensive experiments on SIF and standard benchmarks demon- strate that our method achieves state-of-the-art localization performance and runtime efficiency

    CONCLUSION In this paper, we introduce SIF, the first high–resolution se- mantic inpainting forgery dataset and EfficientIML with a lightweight three-stage backbone of linear complexity and a multi-scale decoder under hierarchical loss supervision. Ex- tensive experiments on S...

  2. [10]

    High-resolution image synthesis with latent diffu- sion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer, “High-resolution image synthesis with latent diffu- sion models,” 2021

    2021

  3. [11]

    CASIA image tampering de- tection evaluation database,

    Jing Dong, Wei Wang, and Tieniu Tan, “CASIA image tampering de- tection evaluation database,” in2013 IEEE China Summit and Inter- national Conference on Signal and Information Processing. July 2013, IEEE

    2013

  4. [12]

    Hybrid image-retrieval method for image-splicing validation,

    Nam Thanh Pham, Jong-Weon Lee, Goo-Rak Kwon, and Chun-Su Park, “Hybrid image-retrieval method for image-splicing validation,” Symmetry, vol. 11, no. 1, pp. 83, 2019

    2019

  5. [13]

    Detecting image splicing using geome- try invariants and camera characteristics consistency,

    Y .-F. Hsu and S.-F. Chang, “Detecting image splicing using geome- try invariants and camera characteristics consistency,” inInternational Conference on Multimedia and Expo, 2006

    2006

  6. [14]

    Mfc datasets: Large-scale benchmark datasets for media forensic challenge evaluation,

    Haiying Guan, Mark Kozak, Eric Robertson, Yooyoung Lee, Amy Yates, Andrew Delgado, Daniel Zhou, Timoth ´ee Kheyrkhah, Jeff Smith, and Jonathan, “Mfc datasets: Large-scale benchmark datasets for media forensic challenge evaluation,” inConference Proceedings. 2019-01-11 00:01:00...

    2019

  7. [15]

    DEFACTO: image and face manipulation dataset,

    Ga ¨el Mahfoudi, Badr Tajini, Florent Retraint, Fr’ed’eric Morain- Nicolier, Jean Luc Dugelay, and Marc Pic”, “DEFACTO: image and face manipulation dataset,” in27th European Signal Processing Con- ference (EUSIPCO 2019), A Coru˜na, Spain”, sep 2019

    2019

  8. [16]

    Coverage – a novel database for copy-move forgery detection,

    Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xu- anjing Shen, and Stefan Winkler, “Coverage – a novel database for copy-move forgery detection,” inIEEE International Conference on Image processing (ICIP), 2016, pp. 161–165

    2016

  9. [17]

    Imd2020: A large-scale annotated dataset tailored for detecting manipulated im- ages,

    Adam Novozamsky, Babak Mahdian, and Stanislav Saic, “Imd2020: A large-scale annotated dataset tailored for detecting manipulated im- ages,” in2020 IEEE Winter Applications of Computer Vision Work- shops (WACVW), March 2020, pp. 71–80

    2020

  10. [18]

    A new bench- mark and model for challenging image manipulation detection,

    Zhenfei Zhang, Mingyang Li, and Ming-Ching Chang, “A new bench- mark and model for challenging image manipulation detection,”Pro- ceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 7, pp. 7405–7413, Mar. 2024

    2024

  11. [19]

    A Semantically Im- pactful Image Manipulation Dataset: Characterizing Image Manipula- tions Using Semantic Significance ,

    Yuwei Chen, Ming-Ching Chang, Mattias Kirchner, Zhenfei Zhang, Xin Li, Arslan Basharat, and Anthony Hoogs, “ A Semantically Im- pactful Image Manipulation Dataset: Characterizing Image Manipula- tions Using Semantic Significance ,” in2025 IEEE/CVF Winter Con- ference on Applic...

    2025

  12. [20]

    Iml-vit: Benchmarking image manipulation localization by vision transformer,

    Xiaochen Ma, Bo Du, Zhuohang Jiang, Ahmed Y . Al Hammadi, and Jizhe Zhou, “Iml-vit: Benchmarking image manipulation localization by vision transformer,” 2023

    2023

  13. [21]

    Mvss-net: Multi-view multi-scale supervised networks for image ma- nipulation detection,

    Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li, “Mvss-net: Multi-view multi-scale supervised networks for image ma- nipulation detection,”IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, pp. 1–14, 2022

    2022

  14. [22]

    Shvit: Single-head vision trans- former with memory efficient macro design,

    Seokju Yun and Youngmin Ro, “Shvit: Single-head vision trans- former with memory efficient macro design,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 5756–5767

    2024

  15. [23]

    Vision- rwkv: Efficient and scalable visual perception with rwkv-like architec- tures,

    Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wenhai Wang, “Vision- rwkv: Efficient and scalable visual perception with rwkv-like architec- tures,”arXiv preprint arXiv:2403.02308, 2024

    2024 arXiv

  16. [24]

    Inpaint anything: Segment anything meets image inpainting,

    Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen, “Inpaint anything: Segment anything meets image inpainting,”arXiv preprint arXiv:2304.06790, 2023

    2023 arXiv

  17. [25]

    Sharegpt4v: Improving large multi-modal models with better captions,

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 370–387

    2024

  18. [26]

    Hierarchical open-vocabulary universal image segmentation,

    Xudong Wang, Shufan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell, “Hierarchical open-vocabulary universal image segmentation,” inThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  19. [27]

    Qwen-vl: A ver- satile vision-language model for understanding, localization, text read- ing, and beyond,

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou, “Qwen-vl: A ver- satile vision-language model for understanding, localization, text read- ing, and beyond,”arXiv preprint arXiv:2308.12966, 2023

    2023 arXiv

  20. [28]

    Pscc-net: Progressive spatio-channel correlation network for image manipulation detection and localization,

    Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu, “Pscc-net: Progressive spatio-channel correlation network for image manipulation detection and localization,”IEEE Transactions on Circuits and Systems for Video Technology, 2022

    2022

  21. [29]

    Tresnet: High performance gpu- dedicated architecture,

    Tal Ridnik, Hussam Lawen, Asaf Noy, Emanuel Ben Baruch, Gi- lad Sharir, and Itamar Friedman, “Tresnet: High performance gpu- dedicated architecture,” 2020

    2020

  22. [30]

    Training data-efficient im- age transformers & distillation through attention,

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv´e J´egou, “Training data-efficient im- age transformers & distillation through attention,” 2021

    2021

  23. [31]

    Decoupled weight decay regular- ization,

    Ilya Loshchilov and Frank Hutter, “Decoupled weight decay regular- ization,” 2019

    2019

  24. [32]

    Sgdr: Stochastic gradient descent with warm restarts,

    Ilya Loshchilov and Frank Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” 2017

    2017

  25. [33]

    Emerging properties in self-supervised vision transformers,

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J ´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the Interna- tional Conference on Computer Vision (ICCV), 2021

    2021

  26. [34]

    Imdl- benco: A comprehensive benchmark and codebase for image manipu- lation detection & localization,

    Xiaochen Ma, Xuekang Zhu, Lei Su, Bo Du, Zhuohang Jiang, Bingkui Tong, Zeyu Lei, Xinyu Yang, Chi-Man Pun, Jiancheng Lv, et al., “Imdl- benco: A comprehensive benchmark and codebase for image manipu- lation detection & localization,”Advances in Neural Information Pro- cessing S...

    2025

  27. [35]

    Cnn-generated images are surprisingly easy to spot...for now,

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros, “Cnn-generated images are surprisingly easy to spot...for now,” inCVPR, 2020

    2020

  28. [36]

    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,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 9535– 9544

    2019

  29. [37]

    Omni-iml: Towards unified image manipulation localization,

    Chenfan Qu, Yiwu Zhong, Fengjun Guo, and Lianwen Jin, “Omni-iml: Towards unified image manipulation localization,” 2024

    2024

  30. [38]

    Evo-vit: Slow-fast token evolution for dynamic vision transformer,

    Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weim- ing Dong, Liqing Zhang, Changsheng Xu, and Xing Sun, “Evo-vit: Slow-fast token evolution for dynamic vision transformer,” inProceed- ings of the AAAI Conference on Artificial Intelligence, 2022, vol. 36, pp. 2964–2972

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.