Pith. sign in

REVIEW 4 major objections 7 minor 44 references

This paper introduces a safety-oriented benchmark showing that current AI-image detectors, and their explanations, fail in high-risk evidentiary scenarios.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 08:47 UTC pith:CFUXE4EQ

load-bearing objection SafeIMG is a useful new benchmark with a credible accuracy gap; the explanation-alignment headline is built on softer evidence than the rest of the paper. the 4 major comments →

arxiv 2607.22745 v1 pith:CFUXE4EQ submitted 2026-07-23 cs.CV cs.AI

AI-generated Images Challenge Visual Trust in High-risk Scenarios

Yi-Zhi Wang , Yichen Xiao , Linan Yue , Weibo Gao , Yichao Du , Pengfei Fang , Shimin Di , Min-Ling Zhang This is my paper
classification cs.CV cs.AI
keywords AI-generated image detectionsafety-critical benchmarkvisual forensicsexplainabilityvision-language modelssynthetic image detectionhuman-AI alignmentcommonsense reasoning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that in public- and individual-safety settings, the real question about an image is not simply whether it is AI-generated but whether its visible content is coherent with the world it claims to depict—and that current detectors do not answer that question. To make the case, the authors built SafeIMG, a benchmark of 1,131 images generated by a state-of-the-art text-to-image system across 12 high-risk scenarios, with human annotations that localize suspicious regions and explain local artefacts, commonsense conflicts, and physical inconsistencies. Their evaluations find that the best vision-language model identifies 49.5% of synthetic images, the best specialized detector 33.1%, and human evaluators 81.7%. When models do produce rationales, they cover only 29.8% of human-annotated anomalies, and coverage drops to 15.0% for commonsense conflicts and 12.0% for physical inconsistencies. These findings, if right, imply that automated detectors are not yet reliable arbiters of visual evidence in high-stakes contexts, and that detection benchmarks should measure whether a decision is grounded in the evidence a human would use.

Core claim

The paper's central claim is that reliable detection of AI-generated images in high-risk scenarios requires more than classifying real versus synthetic; it requires spotting the specific anomalies that make an image implausible as evidence. SafeIMG operationalizes this by collecting 2,652 human-annotated region-level anomalies across 1,131 images generated by a frontier text-to-image model, spanning 12 public- and individual-safety categories, and measuring both detection recall and explanation coverage. On this benchmark the strongest vision-language model reaches 49.5% recall and the strongest specialized detector 33.1%, while human evaluators reach 81.7%. Model rationales hit only 29.8% o

What carries the argument

SafeIMG's evidence-level annotation schema is the load-bearing instrument. For each of the 993 images in the annotated subset, human annotators draw a bounding box around each suspicious region, assign one of seven artefact labels (text, faces, hands, lighting, other, commonsense conflict, physics violation), and write a concise explanation of the anomaly. Detection performance is measured as recall on the synthetic set; explanation alignment is measured by a coverage metric—the proportion of human-annotated anomalies that a model's rationales are judged to hit, using a language-model judge for semantic matching. The benchmark also includes a 138-image 'extremely difficult' subset of images

Load-bearing premise

The headline explanatory-alignment numbers assume the human annotators' region labels are complete and correct and that the language-model judge reliably decides when a model rationale 'hits' a human annotation; if either is noisy or biased, the 29.8%, 15.0%, and 12.0% coverage figures are not robust.

What would settle it

Re-run the alignment analysis on the SafeIMG annotated subset with two or three independent judge models and compute inter-annotator agreement on a re-annotation sample; if the judges disagree substantially on hit/miss, or if independent annotators produce very different bounding boxes and labels, the claimed coverage rates would not be a stable property of the models.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gap is representative, automated detection should not be trusted as the sole gatekeeper for visual evidence in news verification, financial disputes, identity checks, or emergency-response contexts.
  • Detection benchmarks for safety-critical imagery should report scenario-level performance and explanation coverage, not just aggregate accuracy, because category-level differences in SafeIMG are large.
  • Improving detection will require scene-level reasoning about commonsense and physical plausibility, not just better low-level forensics, since models already capture local defects but miss the anomalies humans most often annotate.
  • Robustness to compression and re-encoding should be a standard evaluation axis; models that rely on high-frequency traces lose most of their accuracy under dissemination-like degradation.
  • The hard subset shows that even humans cannot always provide explainable evidence, so some fraction of synthetic images may only be detectable through non-human cues or provenance metadata.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The explanation-coverage metric could double as a training signal: an image detector rewarded for explaining human-annotated anomalies might generalize beyond local artifacts, a direction the paper does not pursue.
  • Because the benchmark uses a single generator, the 49.5%-versus-81.7% gap may not transfer to other generators; an immediate extension is to regenerate or augment SafeIMG with other text-to-image models and compare, which would test how much of the gap is generator-specific.
  • The double-digit hard-image rates in some individual-safety categories, such as transaction records, suggest that as generators improve, text and interface categories may quickly become the hardest rather than the easiest for both humans and models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper introduces SafeIMG, a benchmark of 1,131 AI-generated images produced with GPT Image 2 across 12 public- and individual-safety scenarios, together with 2,652 human region-level annotations that localize suspicious regions and label local artefacts as well as commonsense/physical inconsistencies. The authors evaluate general-purpose VLMs, specialized synthetic-image detectors, and human evaluators on image-level detection, explanation-level alignment, hard images, and robustness to propagation-like degradation. Headline numbers are: best VLM 49.5% AI-image accuracy, best specialized detector 33.1%, human evaluators 81.7%, and model explanations covering only 29.8% of human-annotated anomalies (15.0% for commonsense conflicts, 12.0% for physical inconsistencies). The paper concludes that current detectors lack the accuracy, explanatory alignment, and robustness needed for safe deployment in high-risk evidentiary scenarios.

Significance. If the reported results hold, SafeIMG is a useful contribution: it moves AI-generated-image evaluation from generic realism toward safety-relevant evidentiary contexts, and its region-level human annotations support diagnostic evaluation beyond image-level labels. The public release of data and code is valuable, and the category-level breakdowns in Table 2 give a concrete picture of where current systems fail. The claimed human–model gap is large and policy-relevant. However, the paper's broad conclusions—especially about 'model explanations' and 'current detectors' in general—are currently supported only partially: the alignment analysis rests on one unvalidated LLM-judge pipeline applied to one VLM, and the robustness analysis covers only four VLMs. These gaps are fixable, but they are load-bearing for the abstract's central claim.

major comments (4)
  1. [§3.5, Table 3] The explanatory-alignment headline ('Model explanations cover only 29.8%...') is computed from one single pipeline: Claude Opus 4.6 generates rationales and GPT-5.5 judges whether each rationale 'hits' a human annotation. No prompt for the judge, no validation of the judge against human raters, no inter-annotator reliability for the human region annotations, and no precision or false-positive control are reported. Because a model can emit many rationales, coverage can be inflated by generic statements; because matching is semantic text matching, a rationale mentioning the same artefact type elsewhere in the image may be counted as a hit without region-level verification. The abstract generalizes these numbers to 'model explanations' although no other evaluated VLM is subjected to the alignment analysis. To make this pillar load-bearing, the authors should report judge–human agreement (e.
  2. [§3.4, Figure 6a] The robustness conclusion in the abstract ('current detectors lack ... robustness') is supported by only four VLMs; no specialized detector is evaluated under the degradation pipeline. Moreover, the perturbation protocol is not specified beyond '0x/1x/2x/3x'—it is not stated which JPEG quality factors, resizing factors, or re-encoding steps were used, nor whether the same seeds/transforms were applied across models. Please report the exact transformations, include at least the top specialized detector (FreDect), and provide confidence intervals or repeated-run variability for the large relative drops (e.g., Claude Opus 4.6 falling from 32.6% to 2.9%). As written, the robustness pillar is narrower than the claim.
  3. [§2.4, §3.1] All synthetic images in SafeIMG come from a single generator (GPT Image 2). Since specialized detectors such as CNNSpot and LNP are known to be generator-sensitive, the very low accuracies (7.3% and 9.1%) may largely reflect generator shift rather than the safety-oriented content of the benchmark. The broad statements in the abstract and conclusion about 'current detectors' should either be restricted to the evaluated generator or accompanied by a cross-generator evaluation (even a small one). This does not invalidate the benchmark, but it is necessary for the external validity of the central claim.
  4. [§3.6, Figure 1c; §2.5] The hard subset contains very small per-category sample sizes (e.g., P8 n=2, I1 n=9, I4 n=13), yet the text reports relative reductions of 41.1% and 44.2% and concludes that 'current detection capabilities remain category-dependent.' These estimates have wide confidence intervals and are not statistically grounded. In addition, the hard/annotated triage is based on annotators' inability to find a defensible anomaly, but no inter-annotator agreement (e.g., Cohen's kappa) is reported for the triage or for the region annotations, and the human evaluation uses only three evaluators with no reported variance. Please add reliability statistics and confidence intervals, and temper per-category hard-subset conclusions where n is tiny.
minor comments (7)
  1. [Abstract] The phrase 'specialiseddetector' is missing a space in the full-text version; the abstract should read 'specialized detector'.
  2. [§2.3, Eq. (2)] After 'specificity: The prompt expansion rules U ...' the text appears truncated or malformed; please fix the formatting and complete the sentence.
  3. [§3.1, Eq. (12)] Eq. (12) defines recall for the AI-generated class, not balanced accuracy. The term 'AI detection rate' is fine, but please use it consistently and avoid calling this quantity 'accuracy' without qualification, since the benchmark set contains no real images.
  4. [Figure 1b] The category labels are garbled (e.g., 'esnesmommoC' instead of 'Commonsense'). Please correct the figure labels.
  5. [§3.4, Figure 6a] Specify the exact degradation pipeline: which JPEG quality levels, resizing/re-encoding operations, and whether the perturbations were applied identically to all models. This is needed for reproducibility.
  6. [§3.5, Table 3] The metrics 'Human Covered by AI' and 'AI Supported by Human' are used without formal definitions; please define them precisely and state the unit of aggregation (annotations vs. images).
  7. [§3.3, Figure 5] The trade-off figure is informative, but the real-image accuracy is computed on a separate D_real set; please clarify whether the same prompt/decision rule was used for real and synthetic images, and whether the human evaluators also saw real images.

Circularity Check

0 steps flagged

No significant circularity: SafeIMG results are direct empirical measurements against external human annotations and real-image references, with no fitted parameter or definitional identity between inputs and claims.

full rationale

The paper is an empirical benchmark paper, not a derivation. The construction pipeline (Eqs. 1-11) defines a taxonomy, scenario spaces, prompts, generation, and human annotation; each stage is an operational data-construction step rather than a claim derived from the final results. The headline detection figures (49.5% VLM, 33.1% specialized detector, 81.7% human) are direct measurements using Eq. 12 on SafeIMG, with human labels and real-image sets acting as external references; no parameter is fitted to the test data and then renamed as a prediction. The explanatory-alignment numbers in §3.5 and Table 3 are computed by comparing Claude Opus 4.6 rationales against independent human region annotations, using GPT-5.5 as a semantic judge. This introduces a validation/robustness risk, but it is not circular: the human annotations are external ground truth, the judge is not constructed from the reported coverage values, and the headline accuracy numbers do not reduce to the judge. There are no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in via prior author work. The skeptical concerns about judge validation and inter-annotator reliability are legitimate correctness risks, but they are not instances of the paper's output being equivalent to its input by definition or by fit.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

No free parameters or invented entities; the paper's contribution is a dataset and evaluation protocol. Its main external inputs are human annotations, one image generator, and LLM-based semantic matching.

axioms (3)
  • domain assumption Human annotations define the ground truth for what counts as an anomaly (bounding box, artifact type, explanation).
    Section 2.5 Stage 2: every annotation is a tuple of bounding box, artifact label, and explanation; all coverage metrics in §3.5 are measured against these annotations.
  • domain assumption GPT Image 2 is treated as representative of current frontier image generators for the benchmark's purpose.
    Section 2.4 and all experiments use only GPT Image 2; the abstract and conclusion generalize to 'AI-generated images' broadly.
  • domain assumption GPT-5.5 as a judge yields valid semantic matching between model rationales and human explanations.
    Section 3.5: 'We use GPT-5.5 as the judge model'; no validation of judge agreement against human raters is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 33570 in / 10700 out tokens · 112118 ms · 2026-08-01T08:47:57.143080+00:00 · methodology

0 comments
read the original abstract

Rapid advances in image generation are eroding the evidentiary value of visual content in settings where authenticity can affect public safety and personal reputation. Yet existing detection benchmarks rarely examine synthetic images in public- and individual-safety contexts, where misleading visual content may carry substantial risks. Here we introduce SafeIMG, a safety-oriented benchmark spanning 12 public- and individual-safety scenarios generated using GPT Image 2. Unlike benchmarks centred on generic imagery and image-level labels, SafeIMG evaluates not only whether detectors recognise synthetic images, but also whether their decisions reflect human-identified anomalies. To this end, SafeIMG provides human annotations that localise suspicious regions and explain local artefacts and higher-level commonsense or physical inconsistencies. We evaluate specialized synthetic-image detectors and vision-language models (VLMs), and find that neither provides reliable detection. The strongest VLM identifies only 49.5% of generated images, whereas the best specialised detector identifies 33.1%, compared with 81.7% accuracy for human evaluators. Model explanations cover only 29.8\% of human-annotated anomalies and predominantly capture local defects in text, faces and hands. Their coverage falls to 15.0% for commonsense conflicts and 12.0% for physical inconsistencies, while detection performance deteriorates further after dissemination-induced image degradation. These findings show that current detectors lack the accuracy, explanatory alignment and robustness needed to evaluate AI-generated images reliably across public- and individual-safety settings.

Figures

Figures reproduced from arXiv: 2607.22745 by Linan Yue, Min-Ling Zhang, Pengfei Fang, Shimin Di, WeiBo Gao, Yichao Du, Yichen Xiao, Yi-Zhi Wang.

Figure 1
Figure 1. Figure 1: Overview of SafeIMG and key empirical findings. The figure highlights the challenges of detecting AI￾generated images in high-risk scenarios, where current models exhibit uneven capability and limited understanding of human-perceived artifacts. Corresponding author(s): Linan Yue, Email: lnyue@seu.edu.cn arXiv:2607.22745v1 [cs.CV] 23 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview and distinguishing characteristics of SafeIMG. SafeIMG organises 12 safety-oriented scenarios into public- and individual-safety domains and uses structured prompts to generate realistic, context-rich images with GPT Image 2. Human annotators localise suspicious regions and provide artefact categories and textual explanations, covering both local visual defects and higher-level commonsense or phys… view at source ↗
Figure 3
Figure 3. Figure 3: Construction pipeline of the SafeIMG benchmark for high-risk visual evidence scenarios. SafeIMG is built through five stages. (1) The taxonomy in stage 1 defines 12 high-risk image categories under public or individual safety. (2) For each category, the scenario space specifies the content list ℰ, media form list ℳ, and risk summary R; (3) complete prompts are then constructed before (4) generating images … view at source ↗
Figure 4
Figure 4. Figure 4: SafeIMG evaluation data. (a) Category composition of the 1,131 synthetic images. Bars show the percentages assigned to the annotated (pale gold) and hard (green) subsets. Values give the corresponding image counts and hard-image rates. (b) Within-category distributions of the seven region-level anomaly labels. Cell values are rounded percentages, unrounded rows sum to 100%. (c,d) Explanation-length summari… view at source ↗
Figure 5
Figure 5. Figure 5: Trade-off between AI-image accuracy and real-image accuracy across VLMs on SafeIMG. Points above the diagonal indicate higher AI than real accuracy, and vice versa. and crowds (e.g., in P6, P7, and P8) also increase the difficulty of manual recognition. Overall, the differences in performance across scenarios such as public safety images, private receipts, chat records, and identity endorsement further dem… view at source ↗
Figure 6
Figure 6. Figure 6: (a). AI-image detection accuracy of representative VLMs under network propagation perturbations. (b). Distribution of model-generated rationales that are not supported by human annotations. at 49.5%, while maintaining a moderate real-image accuracy of 60.5%. Similarly, Doubao Seed 2.0 Pro and Gemini 3.1 Pro achieve AI-image accuracies of 38.2% and 39.5%, with real-image accuracies of 65.5% and 78.7%, respe… view at source ↗
Figure 7
Figure 7. Figure 7: Representative cases comparing human evaluations with VLMs. The examples illustrate agreements and disagreements between human and model judgments, showing that humans may be misled by realistic synthetic images, while the model may either mistake imperfections in real images for synthetic artefacts or overlook localized anomalies in AI-generated images. correctly classified it as real, but Claude misclass… view at source ↗
Figure 8
Figure 8. Figure 8: Prevalence difference of artifact labels between false negatives and all AI-generated images. Individual, Public, and Overall denote individual-safety, public-safety, and full-dataset aggregates. Crosses “×” indicate labels absent in both compared sets. Appendix A. Detection Error Analysis on Base VLMs To analyze the failure patterns of base VLMs, we compare the prevalence of each human-annotated artifact … view at source ↗
Figure 9
Figure 9. Figure 9: Case study on P1 - Natural Disasters. The broadcast depicts a severe Harbin blizzard in July, an implausible seasonal event, and reports the temperature range in reversed order. A malicious actor could use such fabricated news imagery to exaggerate a disaster, provoke public panic, or distort travel and emergency-response decisions. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Case study on P2 - Social Unrest, Public Violence, and Attack Incidents. The scene lacks a coherent collision history: the vehicle has no visible impact counterpart, a streetlight is unsupported, and the crowd remains panicked despite police control. A malicious actor could present it as a vehicle attack to incite fear, inflame tensions, or trigger an unwarranted security response [PITH_FULL_IMAGE:figure… view at source ↗
Figure 11
Figure 11. Figure 11: Case study on P3 - Transportation Accidents. The report attributes the crash to tire failure although the visible damage is concentrated on the vehicle body, and the truck’s direction of travel conflicts with the road sign. A malicious actor could use the image to fabricate an accident narrative, falsely assign responsibility, or mislead emergency and traffic-management decisions [PITH_FULL_IMAGE:figures… view at source ↗
Figure 12
Figure 12. Figure 12: Case study on P4 - Fires, Explosions, and Energy Facility Accidents. Bystanders stand inside the police cordon, while the elevated firefighter is too far from the fire to intervene, contradicting incident-response logic. A malicious actor could present it as a major urban fire to cause panic, misrepresent responders, or divert emergency resources. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Case study on P5 - Pollution and Hazardous Material Accidents. The overhead sign and roadside arrow direct traffic in opposite directions, an implausible arrangement during a coordinated hazardous-material response. A malicious actor could use the image to fabricate a chemical spill, provoke evacuation or traffic disruption, or trigger unwarranted public and regulatory action [PITH_FULL_IMAGE:figures/ful… view at source ↗
Figure 14
Figure 14. Figure 14: Case study on P6 - Building and Civil Infrastructure Accidents. The sinkhole scene violates road-safety logic: one side remains accessible without a cordon, and a traffic light stands in the roadway. A malicious actor could circulate it as false evidence of infrastructure failure to create panic, disrupt travel, or manipulate liability claims [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Case study on P7 - Crowd Gathering and Venue Safety Accidents. The nighttime scene conflicts with the stated time and season in Xinjiang, while the schedule repeats one program as both current and upcoming. A malicious actor could use it to fabricate a venue emergency, exaggerate crowd-control failures, or inflame public disorder. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Case study on P8 - Public Health and Biosecurity Events. The triage setup is operationally implausible: the desk sits outside the tent, the thermometer is too small to function as shown, and the notice repeats its date. A malicious actor could use it to invent an outbreak response, spread health misinformation, or undermine trust in public-health institutions [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Case study on I1 - Personal Accidents and Emergencies. The post-fire kitchen is causally inconsistent: power remains on, while heat-sensitive bottles, bags, and food survive the surrounding fire damage. A malicious actor could use it as false evidence of an emergency to solicit money, support insurance fraud, or shift liability [PITH_FULL_IMAGE:figures/full_fig_p030_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Case study on I2 - Private Receipts and Transaction Records. The banking record violates the interface’s expected chronology, placing earlier transactions above later ones in a supposedly ordered ledger. A malicious actor could use the fabricated record to claim a payment or refund that never occurred, enabling fraud, false reimbursement, or manipulation of a financial dispute. 30 [PITH_FULL_IMAGE:figure… view at source ↗
Figure 19
Figure 19. Figure 19: Case study on I3 - Personal Chat and Communication Records. The purported bank-message screenshot conflicts with the institution’s real identity and normal messaging conventions: the logo is incorrect, and timestamps appear in an implausible location. A malicious actor could use it to impersonate a bank, fabricate account activity, or lend credibility to phishing and payment scams [PITH_FULL_IMAGE:figure… view at source ↗
Figure 20
Figure 20. Figure 20: Case study on I4 - Fabricated Scene Evidence and Identity Endorsement. The displayed corporate timeline is not chronological, creating a temporal inconsistency that undermines the claimed business setting. A malicious actor could use the image to fabricate affiliation or endorsement, support investment fraud, or damage reputations. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_20.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 16 linked inside Pith

  1. [1]

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle =

  2. [2]

    Large Scale

    Brock, Andrew and Donahue, Jeff and Simonyan, Karen , booktitle =. Large Scale. 2019 , url =

    2019

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    A Style-Based Generator Architecture for Generative Adversarial Networks , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  4. [4]

    Advances in Neural Information Processing Systems , volume =

    Denoising Diffusion Probabilistic Models , author =. Advances in Neural Information Processing Systems , volume =

  5. [5]

    Diffusion Models Beat

    Dhariwal, Prafulla and Nichol, Alexander , booktitle =. Diffusion Models Beat

  6. [6]

    2112.10741 , archivePrefix =

    Nichol, Alex and Dhariwal, Prafulla and Ramesh, Aditya and Shyam, Pranav and Mishkin, Pamela and McGrew, Bob and Sutskever, Ilya and Chen, Mark , year =. 2112.10741 , archivePrefix =

    Pith/arXiv arXiv

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Vector Quantized Diffusion Model for Text-to-Image Synthesis , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    High-Resolution Image Synthesis with Latent Diffusion Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  9. [9]

    Advances in Neural Information Processing Systems , pages =

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding , author =. Advances in Neural Information Processing Systems , pages =

  10. [10]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Adding Conditional Control to Text-to-Image Diffusion Models , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  11. [11]

    2311.03054 , archivePrefix =

    Tuo, Yuxiang and Xiang, Wangmeng and He, Jun-Yan and Geng, Yifeng and Xie, Xuansong , year =. 2311.03054 , archivePrefix =

    Pith/arXiv arXiv

  12. [12]

    Neural Computing and Applications , volume=

    MEDIC: a multi-task learning dataset for disaster image classification , author=. Neural Computing and Applications , volume=. 2023 , publisher=

    2023

  13. [13]

    Proceedings of the 30th annual ACM symposium on user interface software and technology , pages=

    Rico: A mobile app dataset for building data-driven design applications , author=. Proceedings of the 30th annual ACM symposium on user interface software and technology , pages=

  14. [14]

    arXiv preprint arXiv:2103.14470 , year=

    Spatial dual-modality graph reasoning for key information extraction , author=. arXiv preprint arXiv:2103.14470 , year=

    Pith/arXiv arXiv

  15. [15]

    California Law Review , volume =

    Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security , author =. California Law Review , volume =. 2019 , url =

    2019

  16. [16]

    Social Media + Society , volume =

    Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News , author =. Social Media + Society , volume =. 2020 , doi =

    2020

  17. [17]

    and Farid, Hany , journal =

    Nightingale, Sophie J. and Farid, Hany , journal =. 2022 , doi =

    2022

  18. [18]

    , booktitle =

    Wang, Sheng-Yu and Wang, Oliver and Zhang, Richard and Owens, Andrew and Efros, Alexei A. , booktitle =

  19. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Towards Universal Fake Image Detectors that Generalize Across Generative Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  20. [20]

    2210.06998 , archivePrefix =

    Sha, Zeyang and Li, Zheng and Yu, Ning and Zhang, Yang , year =. 2210.06998 , archivePrefix =

    Pith/arXiv arXiv

  21. [21]

    2303.09295 , archivePrefix =

    Wang, Zhendong and Bao, Jianmin and Zhou, Wengang and Wang, Weilun and Hu, Hezhen and Chen, Hong and Li, Houqiang , year =. 2303.09295 , archivePrefix =

    Pith/arXiv arXiv

  22. [22]

    2306.08571 , archivePrefix =

    Zhu, Mingjian and Chen, Hanting and Yan, Qiangyu and Huang, Xudong and Lin, Guanyu and Li, Wei and Tu, Zhijun and Hu, Hailin and Hu, Jie and Wang, Yunhe , year =. 2306.08571 , archivePrefix =

    Pith/arXiv arXiv

  23. [23]

    2311.12397 , archivePrefix =

    Zhong, Nan and Xu, Yiran and Li, Sheng and Qian, Zhenxing and Zhang, Xinpeng , year =. 2311.12397 , archivePrefix =

    Pith/arXiv arXiv

  24. [24]

    Rethinking the Up-Sampling Operations in

    Tan, Chuangchuang and Liu, Huan and Zhao, Yao and Wei, Shikui and Gu, Guanghua and Liu, Ping and Wei, Yunchao , booktitle =. Rethinking the Up-Sampling Operations in. 2024 , eprint =

    2024

  25. [25]

    A Sanity Check for

    Yan, Shilin and Li, Ouxiang and Cai, Jiayin and Hao, Yanbin and Jiang, Xiaolong and Hu, Yao and Xie, Weidi , year =. A Sanity Check for. 2406.19435 , archivePrefix =

    Pith/arXiv arXiv

  26. [26]

    2024 , eprint =

    Community Forensics: Using Thousands of Generators to Train Fake Image Detectors , author =. 2024 , eprint =

    2024

  27. [27]

    2504.20865 , archivePrefix =

    Pellegrini, Lorenzo and Cozzolino, Davide and Pandolfini, Serafino and Maltoni, Davide and Ferrara, Matteo and Verdoliva, Luisa and Prati, Marco and Ramilli, Marco , year =. 2504.20865 , archivePrefix =

    arXiv

  28. [28]

    Detecting

    Li, Meien and Stamp, Mark , year =. Detecting. 2504.07078 , archivePrefix =

    Pith/arXiv arXiv

  29. [29]

    2504.11015 , archivePrefix =

    Zhu, Chenyang and Zhang, Xing and Sun, Yuyang and Chang, Ching-Chun and Echizen, Isao , year =. 2504.11015 , archivePrefix =

    Pith/arXiv arXiv

  30. [30]

    2509.19841 , archivePrefix =

    Huang, Tai-Ming and Lin, Wei-Tung and Hua, Kai-Lung and Cheng, Wen-Huang and Yamagishi, Junichi and Chen, Jun-Cheng , year =. 2509.19841 , archivePrefix =

    arXiv

  31. [31]

    and Wang, Kai and Yang, Tianyu and Singh, Harsh and Tian, Yapeng , year =

    Yang, Michael and Deng, Shijian and Doan, William T. and Wang, Kai and Yang, Tianyu and Singh, Harsh and Tian, Yapeng , year =. Explainable. 2511.12363 , archivePrefix =

    arXiv

  32. [32]

    2406.16531 , archivePrefix =

    Chen, Yirui and Huang, Xudong and Zhang, Quan and Li, Wei and Zhu, Mingjian and Yan, Qiangyu and Li, Simiao and Chen, Hanting and Hu, Hailin and Yang, Jie and Liu, Wei and Hu, Jie , year =. 2406.16531 , archivePrefix =

    Pith/arXiv arXiv

  33. [33]

    2604.08211 , archivePrefix =

    Hu, You and Zhao, Chenzhuo and Mo, Changfa and Liu, Haotian and Li, Xiaobai , year =. 2604.08211 , archivePrefix =

    Pith/arXiv arXiv

  34. [34]

    2024 , eprint =

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author =. 2024 , eprint =

    2024

  35. [35]

    2403.04692 , archivePrefix =

    Chen, Junsong and Ge, Chongjian and Xie, Enze and Wu, Yue and Yao, Lewei and Ren, Xiaozhe and Wang, Zhongdao and Luo, Ping and Lu, Huchuan and Li, Zhenguo , year =. 2403.04692 , archivePrefix =

    Pith/arXiv arXiv

  36. [36]

    2504.02782 , archivePrefix =

    Yan, Zhiyuan and Ye, Junyan and Li, Weijia and Huang, Zilong and Yuan, Shenghai and He, Xiangyang and Lin, Kaiqing and He, Jun and He, Conghui and Yuan, Li , year =. 2504.02782 , archivePrefix =

    Pith/arXiv arXiv

  37. [37]

    An Empirical Study of

    Chen, Sixiang and Bai, Jinbin and Zhao, Zhuoran and Ye, Tian and Shi, Qingyu and Zhou, Donghao and Chai, Wenhao and Lin, Xin and Wu, Jianzong and Tang, Chao and Xu, Shilin and Zhang, Tao and Yuan, Haobo and Zhou, Yikang and Chow, Wei and Li, Linfeng and Li, Xiangtai and Zhu, Lei and Qi, Lu , year =. An Empirical Study of. 2504.05979 , archivePrefix =

    Pith/arXiv arXiv

  38. [38]

    2025 , howpublished =

    2025

  39. [39]

    2026 , howpublished =

    Introducing. 2026 , howpublished =

    2026

  40. [40]

    2511.21631 , archivePrefix =

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and others , year =. 2511.21631 , archivePrefix =

    Pith/arXiv arXiv

  41. [41]

    International Conference on Learning Representations , year =

    Score-Based Generative Modeling Through Stochastic Differential Equations , author =. International Conference on Learning Representations , year =

  42. [42]

    International Conference on Learning Representations , year =

    Flow Matching for Generative Modeling , author =. International Conference on Learning Representations , year =

  43. [43]

    International conference on machine learning , pages=

    Leveraging frequency analysis for deep fake image recognition , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  44. [44]

    European conference on computer vision , pages=

    Detecting generated images by real images , author=. European conference on computer vision , pages=. 2022 , organization=

    2022