Pith. sign in

REVIEW 2 major objections 4 minor 64 references

SEED: A Benchmark Dataset for Sequential Facial Attribute Editing with Diffusion Models

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper introduces SEED, the first large-scale diffusion-generated facial editing dataset with 91,526 images annotated with ordered edit sequences, and a frequency-aware detector, FAITH, that uses high-frequency cues to recover those…

desk verdict A useful new dataset that currently cannot be trusted on its labels, wrapped in a method claim its own Table 2 contradicts — worth sending out, but only with a demand for major revisions. read the letter →

arxiv 2506.00562 v1 pith:XSQBY6RD submitted 2025-05-31 cs.CV cs.MM

Yule Zhu , Ping Liu , Zhedong Zheng , Wei Liu This is my paper
classification cs.CVcs.MM
keywords facialeditingdatasetsequentialattributediffusionmodelsdetectionfrequency-domainanalysisdiscretewavelettransformfaceforgerybenchmark
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 introduces SEED, a benchmark of 91,526 facial images generated by diffusion models, each edited one to four times in sequence, with labels for the exact attribute sequence, masks, prompts, and editing model used. It argues that this is the first large-scale diffusion-based dataset for sequential facial attribute editing, filling the gap left by the GAN-based SeqDeepFake dataset, which has fewer images and coarser annotations. Alongside the dataset, the paper proposes FAITH, a transformer that injects high-frequency wavelet features into cross-attention to detect and order sequential edits. Experiments report that frequency guidance, especially the DWT high-high sub-band, improves sequence-level accuracy over a spatial-only baseline and over generic deepfake detectors. A sympathetic reader would care because reliable edit-order tracking is needed for provenance, forensics, and evaluation of progressive editing tools, and the dataset makes such research possible at scale.

What carries the argument

The load-bearing mechanism is the frequency-aware cross-attention decoder in FAITH. After a CNN extracts coarse spatial features, a transformer encoder refines them, and the decoder predicts the ordered attribute tokens autoregressively while biasing cross-attention logits by a guidance map built from the high-high (HH) sub-band of a discrete wavelet transform (DWT) of the input image. The HH sub-band captures fine directional textures and edges introduced by diffusion edits, and this explicit high-frequency spatial guidance is what the paper credits for sharper sensitivity to subtle sequential changes.

What would settle it

Take a random sample of SEED images and have independent human raters compare each image against its recorded edit sequence, checking whether each intermediate edit is visibly present and in the recorded order; if a substantial share of samples (say more than 5%) shows missing, wrongly ordered, or proxy-masked edits, the benchmark labels are not trustworthy enough to train or evaluate sequence predictors.

Watch

Extended reading notes

Core claim

SEED demonstrates that large-scale, fine-grained sequential facial editing data can be produced automatically by chaining three diffusion editing pipelines (LEdits, SDXL, and SD3 with UltraEdit) over FFHQ and CelebAMask-HQ source images. Each image is annotated with its exact edit sequence, per-attribute masks, textual prompts, and the editing model used at each step, filtered by quality metrics based on DINO similarity, CLIP similarity, and SSIM. The paper claims this dataset is substantially more realistic, diverse, and finely annotated than existing sequential editing data, and that sequential editing detection should incorporate frequency-domain cues: feeding the DWT HH sub-band as auxiliary spatial guidance into a transformer decoder yields consistent gains in Fixed-, Adaptive-, and Full-Accuracy, with the highest reported average Full-Accuracy of 68.84% on the balanced test partition.

Load-bearing premise

The unstated quality thresholds and the approximate proxy masks (dilated eye region for glasses, upper hair region for hats) jointly carry the guarantee that every retained image truly contains exactly the recorded ordered edits; if those filters mislabel an image, every sequence label, mask, and downstream metric inherits the error.

Editorial extensions

If this is right

  • SEED provides 91,526 labeled images across 200 ordered attribute sequences, so a model can be trained not just to tell edited from unedited faces but to predict the exact ordered list of one to four edits.
  • Because sequence length and editing model are randomized and the recommended partition balances sequence lengths, evaluation on SEED measures performance across difficulty levels rather than on a single easy average.
  • Longer edit sequences sharply reduce all detectors' accuracy, with Full-Accuracy falling from about 93% for one edit to about 24% for four edits, isolating the hardest end of the progressive-editing problem for future research.
  • Frequency-domain guidance, particularly the DWT HH component, improves robustness to JPEG compression and Gaussian noise relative to spatial-only baselines in the reported experiments.

Reading between the lines

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

  • The reported gains of FAITH over the spatial-only baseline are small in absolute terms (about 0.3 points in average Full-Accuracy), so a reader should not over-read the frequency mechanism's advantage without significance testing or confidence intervals.
  • The dataset pipeline could be extended to non-facial subjects or to longer edit chains, and publishing the exact quality thresholds and mask-proxy rules would let others reproduce failure modes and measure label noise.
  • Because the glasses and hat masks are proxies (dilated eye region and upper hair region), images involving additive accessories may carry the most label noise; a human-verified subset would make SEED a cleaner training and evaluation resource.
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

2 major / 4 minor

Summary. The manuscript introduces SEED, a benchmark of 91,526 facial images obtained by applying 1 to 4 sequential diffusion-based edits (LEdits, SDXL, and SD3 fine-tuned with UltraEdit) to faces from FFHQ and CelebAMask-HQ, with per-image annotations of edit order, masks, prompts, and editing method. It also proposes FAITH, a frequency-aware transformer that incorporates DWT high-frequency features in cross-attention for sequential edit prediction, and reports experiments comparing FAITH variants (DCT, FFT, DWT) with SeqFakeFormer and adapted binary detectors. The paper claims that FAITH significantly improves sequential editing detection, particularly for longer sequences, and that SEED provides large-scale, precisely annotated data for this task.

Significance. The dataset resource is potentially valuable: it offers a much larger, diffusion-based sequential editing benchmark than SeqDeepFake, with multiple editing pipelines, per-step masks, prompts, and editing metadata, and the authors commit to releasing data and code. Comparing DCT, FFT, and DWT as frequency cues in a unified architecture is also a useful baseline study. However, the central experimental claim is contradicted by the paper's own Table 2 at the longest sequence length, and the dataset's ground-truth validity is not established because the label-generation pipeline relies on approximate masks and undisclosed global quality thresholds. I do not see circular reasoning in the evaluation protocol; the main concerns are internal consistency of the FAITH results and missing attribute-level validation of the SEED labels.

major comments (2)
  1. [Section 5.2, Table 2] Table 2 contradicts the paper's central claim that frequency-domain features significantly improve detection of longer editing sequences. For sequence length 4, SeqFakeFormer achieves Fixed/Adaptive/Full accuracy of 50.14/50.14/24.55, while all FAITH variants achieve Fixed/Adaptive values below 50 and Full-Acc values of 22.95 (DCT), 23.05 (FFT), and 23.35 (DWT). Thus FAITH is worse than the spatial baseline on every metric at the longest sequence length. The average Full-Acc improvement of DWT over SeqFakeFormer is only 0.31 points (68.84 vs 68.53), and no variance or significance test is reported. The abstract and Section 5.2 claims about enhanced sensitivity to subtle sequential changes therefore need to be substantially qualified or the experimental evidence reconsidered.
  2. [Section 3.1, Preprocessing and Image Quality Evaluation] The SEED ground-truth labels are not validated at the attribute level. The text states that glasses masks are obtained by dilating the eye mask and hat masks by using the upper hair region as a proxy, and that images are retained only if they pass 'predefined quality thresholds' on DINO similarity, CLIP similarity, and SSIM, yet no thresholds or filtering rules are reported. These global similarity metrics do not verify that the intended attribute change actually occurred, so an image in which an edit fails or modifies a different attribute can be retained with an incorrect sequence label. Because every downstream task on SEED is supervised by these labels, the benchmark needs a reported label audit, such as human evaluation on a sample or attribute-classifier verification, before the claimed annotation accuracy can be accepted.
minor comments (4)
  1. [Table 3(a)] The SeqFakeFormer row in Table 3(a) appears misaligned: the nine reported values (78.68, 61.60, 63.67, 44.41, 75.83, 59.24, 62.01, 66.63, 39.33) do not line up with the Fixed/Adaptive/Full columns across the three compression levels and should be corrected.
  2. [References [56]] DINO is cited to the DETR object-detection paper, but if the authors use DINOv1 or DINOv2 self-supervised features for quality assessment, the appropriate reference should be provided.
  3. [Section 1 and Table 1] The text says SeqDeepFake contains fewer than 50K images, while Table 1 reports two size numbers (35,166 and 49,920); please clarify which split or variant corresponds to which column.
  4. [Section 5.2] No error bars, standard deviations, or significance tests are reported; given that the main reported improvement is 0.31 Full-Acc points, a multi-run analysis is needed to establish that the difference is not noise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FAITH is trained and tested on disjoint splits of SEED, and no fitted input is presented as a prediction.

full rationale

The paper's core claims are empirical dataset construction and benchmarking rather than analytical derivations. SEED is generated by applying diffusion editing pipelines, and each image's edit sequence is recorded from the pipeline's own random selection; FAITH is then trained on one partition and evaluated on a held-out partition of SEED using standard cross-entropy supervision and sequence-level accuracy metrics. No equation in the paper equates an output to an input, no fitted parameter is relabeled as a prediction, and no load-bearing uniqueness theorem or self-citation is invoked. The self-references appear only in related-work and application contexts. The undisclosed quality thresholds and approximate masks for glasses and hats are annotation-validity and generalization concerns, not instances of circular reasoning as defined by the reviewer rules. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central load is empirical reliability of automatically generated labels rather than mathematical derivation. The paper assumes mask fidelity, prompt fidelity, and quality-filter fidelity without reporting threshold values or failure analysis. No invented physical or conceptual entities are introduced.

free parameters (1)
  • Quality thresholds for DINO, CLIP, and SSIM = Not disclosed
    Images are retained only if they exceed 'predefined quality thresholds' in Section 3.1. The actual values are never given, and they determine which images enter SEED and therefore affect every downstream result.
assumptions (3)
  • domain assumption CelebAMask-HQ face parsing masks accurately constrain each diffusion edit to the intended facial region, so edits are disentangled and sequence labels are correct.
    Invoked in Section 3.1 Preprocessing. If masks leak or fail, the attribute masks and ordered edit labels do not describe the actual image changes.
  • domain assumption DINO, CLIP, and SSIM thresholds filter edited images without systematically selecting for easier-to-detect artifacts or mislabeled failures.
    Section 3.1 Image Quality Evaluation. Thresholds are omitted and no failure analysis is reported, so the retained set's label quality is assumed.
  • domain assumption LLaVA-generated prompts are semantically aligned with the intended attribute edit, so the prompt drives the correct edit in LEdits, SDXL, or SD3.
    Section 3.1 Preprocessing. Prompt quality is not independently evaluated, but the entire sequence label depends on the prompt producing the intended edit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SEED: A Benchmark Dataset for Sequential Facial Attribute Editing with Diffusion Models." pith.science (2026). https://pith.science/paper/XSQBY6RD

@misc{pith2026250600562,
  author       = {Pith},
  title        = {Pith review of: SEED: A Benchmark Dataset for Sequential Facial Attribute Editing with Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XSQBY6RD}},
  note         = {Machine review of arXiv:2506.00562}
}
read the original abstract

Diffusion models have recently enabled precise and photorealistic facial editing across a wide range of semantic attributes. Beyond single-step modifications, a growing class of applications now demands the ability to analyze and track sequences of progressive edits, such as stepwise changes to hair, makeup, or accessories. However, sequential editing introduces significant challenges in edit attribution and detection robustness, further complicated by the lack of large-scale, finely annotated benchmarks tailored explicitly for this task. We introduce SEED, a large-scale Sequentially Edited facE Dataset constructed via state-of-the-art diffusion models. SEED contains over 90,000 facial images with one to four sequential attribute modifications, generated using diverse diffusion-based editing pipelines (LEdits, SDXL, SD3). Each image is annotated with detailed edit sequences, attribute masks, and prompts, facilitating research on sequential edit tracking, visual provenance analysis, and manipulation robustness assessment. To benchmark this task, we propose FAITH, a frequency-aware transformer-based model that incorporates high-frequency cues to enhance sensitivity to subtle sequential changes. Comprehensive experiments, including systematic comparisons of multiple frequency-domain methods, demonstrate the effectiveness of FAITH and the unique challenges posed by SEED. SEED offers a challenging and flexible resource for studying progressive diffusion-based edits at scale. Dataset and code will be publicly released at: https://github.com/Zeus1037/SEED.

Figures

Figures reproduced from arXiv: 2506.00562 by the authors.

Figure 1
Figure 1. Overview of the SEED Dataset Generation Pipeline. Facial images first undergo mask and prompt generation using [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Samples from SEED demonstrating sequential fa [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed FAITH architecture. The model explicitly integrates high-frequency domain information [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 36 canonical work pages

  1. [1]

    Ahmed, T

    N. Ahmed, T. Natarajan, and K.R. Rao. 1974. Discrete Cosine Transform.IEEE Trans. Comput.C-23, 1 (1974), 90–93. doi:10.1109/T-C.1974.223784

    arXiv 1974

  2. [2]

    Zhongjie Ba, Qingyu Liu, Zhenguang Liu, Shuang Wu, Feng Lin, Li Lu, and Kui Ren. 2024. Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection.Proceedings of the AAAI Conference on Artificial Intelligence38, 2 (Mar. 2024), 719–728. doi:10.1609/aaai.v38i2.27829

    doi:10.1609/aaai.v38i2.27829 2024

  3. [3]

    Jiaxuan Chen, Jieteng Yao, and Li Niu. 2024. A Single Simple Patch is All You Need for AI-generated Image Detection.arXiv:2402.01123(2024)

    arXiv 2024

  4. [4]

    Zhongxi Chen, Ke Sun, Ziyin Zhou, Xianming Lin, Xiaoshuai Sun, Liujuan Cao, and Rongrong Ji. 2024. DiffusionFace: Towards a Comprehensive Dataset for Diffusion-Based Face Forgery Analysis.arXiv preprint arXiv:2403.18471(2024)

    arXiv 2024

  5. [5]

    Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. 2024. Exploiting Style Latent Flows for Generalizing Deepfake Video Detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 1133–1143

    work page 2024

  6. [6]

    Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. 2023. Intriguing Properties of Synthetic Images: From Generative Ad- versarial Networks to Diffusion Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. 973–982

    work page 2023

  7. [7]

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah

  8. [8]

    Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. 2024. FluxSpace: Disentan- gled Semantic Editing in Rectified Flow Transformers. arXiv:2412.09611 [cs.CV]

    arXiv 2024

Show all 64 references
  1. [9]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. InCVPR

    2009

  2. [10]

    Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. 2021. Sharpness-aware Minimization for Efficiently Improving Generalization. InICLR

    2021

  3. [11]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks.Commun. ACM63, 11 (Oct. 2020), 139–144. doi:10.1145/3422622

    2020 doi

  4. [12]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Nets. InAdvances in Neural Information Processing Systems, Z. Ghahra- mani, M. Welling, C. Cortes, N. Lawrence, and...

    2014

  5. [13]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InCVPR

    2016

  6. [14]

    Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. 2021. ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis. InCVPR

    2021

  7. [15]

    Arash Heidari, Nima Jafari Navimipour, Hasan Dag, and Mehmet Unal. 2024. Deepfake detection using deep learning methods: A systematic and compre- hensive review.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery14, 2 (2024), e1520

    2024

  8. [16]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InNeurIPS

    2020

  9. [17]

    Cheng-Yao Hong, Yen-Chi Hsu, and Tyng-Luh Liu. 2024. Contrastive Learning for DeepFake Classification and Localization via Multi-Label Ranking. InCVPR

    2024

  10. [18]

    Bingwen Hu, Zhedong Zheng, Ping Liu, Wankou Yang, and Mingwu Ren. 2021. Unsupervised Eyeglasses Removal in the Wild.IEEE Transactions on Cybernetics 51, 9 (2021), 4373–4385. doi:10.1109/TCYB.2020.2995496

    2021

  11. [19]

    Zhikun Huang, Zhedong Zheng, Chenggang Yan, Hongtao Xie, Yaoqi Sun, Jianzhong Wang, and Jiyong Zhang. 2021. Real-world automatic makeup via identity preservation makeup net. InInternational Joint Conference on Artificial Intelligence. International Joint Conference on Artifici...

    2021

  12. [20]

    Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu

  13. [21]

    Tero Karras, Samuli Laine, and Timo Aila. 2019. A Style-Based Generator Archi- tecture for Generative Adversarial Networks. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 4396–4405. doi:10.1109/CVPR. 2019.00453

    2019

  14. [22]

    Hyunsu Kim, Yunjey Choi, Junho Kim, Sungjoo Yoo, and Youngjung Uh. 2021. Exploiting Spatial Dimensions of Latent in GAN for Real-Time Image Editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 852–861

    2021

  15. [23]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. 2019. An Introduction to Variational Autoencoders.Foundations and Trends®in Machine Learning12, 4 (2019), 307–

    2019

  16. [24]

    Nguyen, Junichi Yamagishi, and Isao Echizen

    Trung-Nghia Le, Huy H. Nguyen, Junichi Yamagishi, and Isao Echizen. 2021. OpenForensics: Large-Scale Challenging Dataset for Multi-Face Forgery Detec- tion and Segmentation In-the-Wild. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 10117–10127

    2021

  17. [25]

    Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. 2020. MaskGAN: Towards Diverse and Interactive Facial Image Manipulation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2020

  18. [26]

    Yunfei Li, Yuezun Li, Xin Wang, Baoyuan Wu, Jiaran Zhou, and Junyu Dong

  19. [27]

    Baoping Liu, Bo Liu, Ming Ding, and Tianqing Zhu. 2024. Detection of Diffusion Model-Generated Faces by Assessing Smoothness and Noise Tolerance. In2024 IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB). 1–6. doi:10.1109/BMSB62888.2024.10608232

    2024

  20. [28]

    Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. 2023. TI2Net: Temporal Identity Inconsistency Network for Deepfake Detection. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 4691–4700

    2023

  21. [29]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruc- tion Tuning

    2023

  22. [30]

    Jiaxin Liu, Jia Wang, Saihui Hou, Min Ren, Huijia Wu, and Zhaofeng He. 2025. Beyond Face Swapping: A Diffusion-Based Digital Human Benchmark for Multi- modal Deepfake Detection.arXiv preprint arXiv:2505.16512(2025)

    2025

  23. [31]

    Ping Liu, Qiqi Tao, and Joey Zhou. 2024. Robust Deepfake Detection by Ad- dressing Generalization and Trustworthiness Challenges: A Short Survey. In Proceedings of the 1st ACM Multimedia Workshop on Multi-modal Misinformation Governance in the Era of Foundation Models. 3–11

    2024

  24. [32]

    Ping Liu, Qiqi Tao, and Joey Tianyi Zhou. 2024. Evolving from Single-modal to Multi-modal Facial Deepfake Detection: A Survey. arXiv:2406.06965 [cs.CV] https://arxiv.org/abs/2406.06965

    2024 arXiv

  25. [33]

    Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. 2023. DF-Platter: Multi-Face Heterogeneous Deepfake Dataset. InCVPR

    2023

  26. [34]

    Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. 2024. LAA-Net: Localized Artifact Attention Network for Quality-Agnostic and Generalizable Deepfake Detection. InProceedings of the IEEE/CVF Conference...

    2024

  27. [35]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  28. [36]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InProceedings ...

    2021

  29. [37]

    Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. 2024. Towards the Detection of Diffusion Model Deepfakes. InProceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, VISIGRAPP 2024, Volume 4: VI...

    2024 doi

  30. [38]

    Roberts and Clifford T

    Richard A. Roberts and Clifford T. Mullis. 1987.Digital signal processing. Addison- Wesley Longman Publishing Co., Inc., USA

    1987

  31. [39]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

    2022

  32. [40]

    Rui Shao, Tianxing Wu, and Ziwei Liu. 2022. Detecting and recovering sequential deepfake manipulation. InEuropean Conference on Computer Vision. Springer, 712–728

    2022

  33. [41]

    Rui Shao, Tianxing Wu, and Ziwei Liu. 2025. Robust Sequential DeepFake Detection.International Journal of Computer Vision133 (01 2025), 3278–3295. doi:10.1007/s11263-024-02339-6

    2025 doi

  34. [42]

    Chao Shuai, Jieming Zhong, Shuang Wu, Feng Lin, Zhibo Wang, Zhongjie Ba, Zhenguang Liu, Lorenzo Cavallaro, and Kui Ren. 2023. Locate and Verify: A Two-Stream Network for Improved Deepfake Detection. InProceedings of the 31st ACM International Conference on Multimedia(Ottawa ON...

    2023

  35. [43]

    Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. 2023. Robustness and generalizability of deepfake detection: A study with diffusion models.arXiv preprint arXiv:2309.02218(2023)

    2023 arXiv

  36. [44]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising Diffusion Implicit Models.arXiv:2010.02502(October 2020). https://arxiv.org/abs/2010. 02502

    2020 arXiv

  37. [45]

    Chuangchuang Tan, Ping Liu, RenShuai Tao, Huan Liu, Yao Zhao, Baoyuan Wu, and Yunchao Wei. 2024. Data-Independent Operator: A Training-Free Artifact Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Yule Zhu, Ping LiuB, Zhedong Zheng, and Wei LiuB Representation Extracto...

    2024 arXiv

  38. [46]

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. 2024. Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning.Proceedings of the AAAI Conference on Artificial Intelligence38, 5 (Mar. 2024), 5052–506...

    2024 doi

  39. [47]

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. 2024. Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2024

  40. [48]

    Linoy Tsaban and Apolinário Passos. 2023. Ledits: Real image editing with ddpm inversion and semantic guidance.arXiv preprint arXiv:2307.00522(2023)

    2023 arXiv

  41. [49]

    Jiacheng Wang, Ping Liu, Jingen Liu, and Wei Xu. 2024. Text-Guided Eyeglasses Manipulation With Spatial Constraints.Trans. Multi.26 (Jan. 2024), 4375–4388. doi:10.1109/TMM.2023.3322326

    2024

  42. [50]

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. 2023. DIRE for Diffusion-Generated Image Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 22445–22455

    2023

  43. [51]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing(2004)

    2004

  44. [52]

    Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. 2022. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models.arXiv preprint arXiv:2210.14896(2022)

    2022 arXiv

  45. [53]

    Ruiyang Xia, Decheng Liu, Jie Li, Lin Yuan, Nannan Wang, and Xinbo Gao

  46. [54]

    Xingxi Yin, Jingfeng Zhang, Zhi Li, Yicheng Li, and Yin Zhang. 2025. InstructAttribute: Fine-grained Object Attributes editing with Instruction. arXiv:2505.00751 [cs.CV]

    2025 arXiv

  47. [55]

    Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. 2025. Learning Natural Consistency Representation for Face Forgery Video Detection. InECCV

    2025

  48. [56]

    Ni, and Heung-Yeung Shum

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. 2023. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. InICLR

    2023

  49. [57]

    Yuxuan Zhang, Lifu Wei, Qing Zhang, Yiren Song, Jiaming Liu, Huaxia Li, Xu Tang, Yao Hu, and Haibo Zhao. 2024. Stable-Makeup: When Real-World Makeup Transfer Meets Diffusion Model.arXiv preprint arXiv:2403.07764(2024)

    2024 arXiv

  50. [58]

    Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. 2024. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. B...

    2024

  51. [59]

    Dragos,-Constantin T, ânt,aru, Elisabeta Oneat,ă, and Dan Oneat,ă. 2024. Weakly- Supervised Deepfake Localization in Diffusion-Generated Images. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 6258–6268. Received 20 February 2007; re...

    2024

  52. [392]

    doi:10.1561/2200000056

    doi

  53. [2021]

    InProceedings of the IEEE/CVF International Conference on Computer Vision

    Talk-to-Edit: Fine-Grained Facial Editing via Dialog. InProceedings of the IEEE/CVF International Conference on Computer Vision. 13799–13808

  54. [2023]

    Diffusion Models in Vision: A Survey.IEEE Transactions on Pattern Analysis and Machine Intelligence(2023)

    2023

  55. [2024]

    Mmnet: multi-collaboration and multi-supervision network for sequential deepfake detection.IEEE Transactions on Information Forensics and Security (2024)

    2024

  56. [2025]

    InProceedings of the Winter Conference on Applications of Computer Vision (W ACV)

    Texture Shape and Order Matter: A New Transformer Design for Sequential DeepFake Detection. InProceedings of the Winter Conference on Applications of Computer Vision (W ACV). 202–211

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.