Pith. sign in

REVIEW 4 major objections 4 minor 91 references

Deployed AI-image detectors can adapt to previously unseen generators by self-training on unlabeled test images through a curriculum, and the paper reports large balanced-accuracy gains across five benchmarks.

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

2026-08-05 00:40 UTC pith:GGCWMA7M

load-bearing objection Competent, well-ablated TTA-for-AIGC paper whose core claim is plausible but whose headline protocol is transductive and whose pseudo-label reliability on truly novel generators is only measured on one benchmark. the 4 major comments →

arxiv 2608.00559 v1 pith:GGCWMA7M submitted 2026-08-01 cs.CV

Test-Time Curriculum for Open-Set AIGC Detection

Yiqian Zhang , Zheyuan Gu , Xiangzhao Hao , Zefeng Zhang , Jingjia Mao , Jiahao Hu , Jiaxu Miao , Jun Yu
show 3 more authors
Zhenyu Zhang Shuohuan Wang Yu Sun
This is my paper
classification cs.CV
keywords test-time adaptationAI-generated image detectiondeepfake detectionpseudo-label self-trainingcurriculum learningcross-scale fusionopen-set detectionAIGCGuard benchmark
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.

The paper's central claim is that AI-generated-image detectors, normally trained once and frozen, can keep improving after deployment by adapting to unlabeled test images. The proposed method, Test-Time Curriculum (TTC), first trusts only the most confident predictions, then gradually relaxes the confidence threshold and retrains the detector on pseudo-labeled samples while keeping a class balance. On the paper's reported numbers, this raises balanced accuracy—the average of real-image and fake-image accuracies—from 81.6% to 93.0% on AIGCDetect, from 82.9% to 99.1% on AIGI-Holmes, from 80.1% to 92.5% on GenImage, and from 62.4% to 91.1% on Chameleon. The practical significance is that deployed detectors could update on unlabeled image streams as new generators appear, without collecting labels or retraining from scratch.

Core claim

The paper sets out to establish that test-time adaptation is a viable generalization paradigm for open-set AI-generated image detection. Instead of a static train-once-and-deploy detector, TTC repeatedly self-trains on the unlabeled target batch: each round selects low-uncertainty predictions as reliable pseudo labels, balances the number of real and fake samples to prevent class-gradient bias, and then chooses the hardest samples within each reliable class as the adaptation set. An L2 anchor to the initial parameters and a label-free safety check guard against drift. Cross-scale pseudo-label refinement produces the pseudo labels by aggregating predictions across three resolutions and multip

What carries the argument

The core mechanism is a two-stage curriculum sampler combined with cross-scale pseudo-label refinement. Reliability is measured by binary entropy, and samples below a round-dependent entropy threshold form the reliable set. Class balance is enforced by taking an equal count from reliable fake and reliable real predictions, and within each class the selected samples are the hardest ones that still satisfy the reliability threshold. The update loss is binary cross-entropy on pseudo labels plus an L2 penalty pulling parameters back toward the pretrained weights. Multi-scale predictions are aggregated with log-sum-exp pooling per scale, a cross-scale mean is used for pseudo-label selection, and

Load-bearing premise

The method assumes that when the base detector is highly confident on an unseen generator's images, it is usually correct; if the detector is systematically and confidently wrong on a new generator family, the curriculum trains on those errors and adaptation can degrade rather than help.

What would settle it

Find a realistic unseen-generator shift in which the base detector is confidently wrong, for example an unseen model whose fake images almost all receive predicted fake-probability above 0.9 while actually being misclassified. If TTC's balanced accuracy after adaptation drops materially below the frozen base detector and the label-free safety check does not roll back, the entropy-reliability premise fails. More directly, measure the accuracy of the selected pseudo labels on a new shift: the paper's oracle analysis predicts they stay near 90%, so observing substantially lower selected-label acc

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If TTC works as reported, an AIGC detector can be updated on a batch of unlabeled deployment images, so data-curation pipelines that already process unlabeled collections can supply adaptation data without extra annotation.
  • The gains transfer beyond the adapted images: evaluation on a disjoint remainder after excluding adaptation samples differs by at most 0.10 BAcc, indicating the adaptation learns generalizable cues rather than memorizing the batch.
  • The same protocol improves other detector backbones, such as AIDE and C2P-CLIP, to levels above the strong static baseline DDA, suggesting the adaptation scheme is architecture-independent.
  • Full test-set access is not required: 10K adaptation images yield 94.46% BAcc in 0.79 GPU hours, close to the 93.59% of full-set adaptation at 11.16 GPU hours, so the method can be used at lower cost.
  • Mixed-domain adaptation over all generators together outperforms generator-specific adaptation on average, because shared artifacts across generators help difficult subsets, though generator-specific adaptation avoids some per-subset negative transfer.

Where Pith is reading between the lines

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

  • Editorial inference: The paper's oracle analysis suggests a 90% pseudo-label accuracy threshold separates stable from degraded adaptation; if that holds across generator families, any future source of more reliable pseudo labels, such as detector ensembles, could push TTC-style methods further with diminishing returns above 90%.
  • Editorial inference: The round-based threshold relaxation maps naturally onto streaming or online adaptation, where batches of unlabeled images arrive over time; the paper explicitly leaves this as future work, but the curriculum schedule appears directly transferable to that setting.
  • Editorial inference: Because the noisy-or output is explicitly not a calibrated posterior, downstream applications that need a probability, such as audit prioritization or risk scoring, would require a separate calibration step on top of TTC.
  • Editorial inference: AIGCGuard's controlled design, with one real image and 40 matched generated counterparts per prompt, could support attribution and generator-identification tasks beyond binary detection, since the generator identity of every fake image is known.

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 / 4 minor

Summary. The paper proposes Test-Time Curriculum (TTC), a test-time adaptation framework for open-set AI-generated-image detection. TTC pseudo-labels unlabeled test images using entropy-based reliable selection, balances the two classes, selects hard-within-reliable samples, anchors updates to the initial detector via L2 regularization, and uses cross-scale pseudo-label refinement plus noisy-or fusion at inference. The authors also introduce AIGCGuard, a new benchmark with 3,100 real images and 124,000 generated images from 40 recent text-to-image models. Experiments on five benchmarks report large balanced-accuracy gains over a Base detector and over prior static AIGC detectors, with extensive ablations, sensitivity analyses, run-to-run stability, and a disjoint adaptation/evaluation check.

Significance. If the empirical claims hold, the paper makes a useful contribution: it reframes open-set AIGC detection as a test-time adaptation problem and provides a model-agnostic recipe that improves deployed detectors on unlabeled target batches. The empirical study is unusually thorough for this area: component ablations (Table 3), oracle-controlled pseudo-label quality analysis (Table 13), disjoint adaptation/evaluation splits (Table 10), run-to-run stability (Table 11), sensitivity to curriculum schedules, patch budgets, LSE temperature, and patch selection, plus generalization to two additional backbones (Table 9). The AIGCGuard benchmark is a valuable asset for evaluating modern text-to-image generators. The main weakness is that the paper's central claim of robust adaptation to truly unseen generators rests on pseudo-label reliability, and the evidence for that reliability is incomplete on the hardest benchmarks. The safety-check mechanism that is supposed to prevent harmful adaptation is also underspecified. These issues are fixable but currently block full confidence in the headline claim.

major comments (4)
  1. [§4.1, Table 10] The primary evaluation protocol is transductive: TTC is adapted on the full unlabeled target split and then evaluated on the same images. The disjoint-split check in Table 10 is performed only on AIGCDetect; the headline numbers for AIGI-Holmes, GenImage, Chameleon, and AIGCGuard all evaluate on the adaptation set. Since the abstract claims a 'practical' framework and improvements under 'unseen-generator settings', the reader needs to know whether gains transfer to non-overlapping images on every benchmark, not just one. Please add non-overlapping evaluation for all five benchmarks, or explicitly state that the main numbers include same-image fitting and restate the claims accordingly.
  2. [§A.8, Table 13; §A.12] The load-bearing assumption is that low-entropy predictions are reliable enough for self-training (Eqs. 2–4). The paper measures in-situ pseudo-label accuracy only on AIGCDetect (Table 13a: 92.6/91.7/92.6%). No such measurement is reported for AIGCGuard or for the AIGI-Holmes subsets where the Base detector is near chance (SD3.5-L 50.5%, FLUX 50.6%). The oracle-controlled experiment (Table 13b) shows a cliff: BAcc drops from 95.1 at 90% pseudo-label accuracy to 79.3 at 80%, and §A.12 explicitly acknowledges that 'systematic high-confidence errors ... may prevent useful adaptation.' This is precisely the failure mode for unseen generators. To support the central claim, report post-hoc pseudo-label accuracy per benchmark (and per challenging subset), and verify that the selected pseudo labels stay in the >90% regime for the hardest shifts. Without this, the 'unseen-generator' claim is not
  3. [§A.9] The label-free safety check is a key mechanism for preventing harmful adaptation, but its thresholds are described only as 'fixed' and are never specified. This makes the method irreproducible and unvalidatable. Moreover, the check uses only prediction statistics (fake prior and average certainty); under systematic confident-wrong predictions—the exact scenario highlighted in §A.12—these statistics may not detect a harmful round. Please provide the exact threshold values and add a controlled experiment that simulates high-confidence incorrect pseudo-labels (e.g., flipping high-confidence fake labels to real, or vice versa) to show whether the rollback triggers and preserves the Base detector.
  4. [§4.1, §A.7, Eq. (2)] There is an inconsistency in the curriculum threshold description. The main text says '3 curriculum rounds with confidence thresholds (0.9, 0.8, 0.7), converted to corresponding binary-entropy thresholds τ'. But binary entropy in natural log has a maximum of ln 2 ≈ 0.693, so τ = 0.9 is impossible; if log base 2 is used, the entropy corresponding to confidence 0.9 is 0.469, not 0.9. The schedules in Fig. 4 and §A.7 (e.g., 0.9, 0.8, 0.7, 0.6, 0.5) appear to be treated as entropy thresholds directly. Please clarify the log base and give the actual τ_r values used in the main experiments.
minor comments (4)
  1. [Fig. 1 caption] Typo: 'After Adaption' should be 'After Adaptation'.
  2. [Table 13] Panel (a) labels the method 'Our'; consider 'TTC' to match the rest of the paper.
  3. [§A.8] The oracle-controlled experiment fixes the adaptation samples and corrupts pseudo-labels with ground truth. Please clarify whether the 95.1% at both 95% and 90% accuracy is a plateau or a typo, and state the number of adaptation samples used.
  4. [§3.3, Eq. (7)] The LSE pooling formula would be easier to parse if the log base and the handling of degenerate patches (Nℓ=0) were stated explicitly.

Circularity Check

0 steps flagged

No significant circularity: the self-training loop is self-referential by design, but the paper provides ground-truth validation and a non-overlapping split check that bound the transductive component.

full rationale

TTC is a pseudo-label self-training method: Eq. (3)-(4) define reliable samples via the detector's own low-entropy predictions, Eq. (6) updates the detector on those pseudo-labels, and the primary evaluation is on the same target batch (Section 4.1: 'The adapted detector is then evaluated on the same target collection, making the primary protocol transductive'). This is a legitimate methodological self-reference rather than a circular derivation, and the paper directly checks whether reported gains are same-sample artifacts. Table 10 and Section A.5 show that evaluating on the non-overlapping remainder after removing adaptation samples changes BAcc by at most 0.10 points, stating the gains 'are not mainly caused by same-sample fitting.' Section A.8 (Table 13) independently measures selected pseudo-label accuracy against ground truth (92.6/91.7/92.6% across rounds) and shows an oracle-controlled cliff below 90% accuracy, so the reliability assumption is empirically examined rather than assumed. The A.12 limitation that 'systematic high-confidence errors... may prevent useful adaptation' is an honest caveat about the entropy-reliability assumption, not a circular step. Self-citations (e.g., ERNIE 5.0, Seedream, Cheers, Ernie-Image) appear only as general context or as generators in the new AIGCGuard benchmark; no load-bearing argument is reduced to an author's own prior theorem or uniqueness claim. Therefore no specific derivation reduces to its own input.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The ledger captures tuned constants (thresholds, temperatures, patch budgets, weights) and the behavioral assumptions the adaptation loop relies on. No new physical or mathematical entities are introduced; AIGCGuard is a benchmark artifact, not a postulated entity.

free parameters (7)
  • Curriculum confidence thresholds (0.9, 0.8, 0.7) converted to entropy thresholds τ_r = 0.9 / 0.8 / 0.7 confidence
    Chosen by hand; Fig 4 shows schedule choice changes final BAcc substantially across schedules (roughly 85.6 to 94.9), so the default schedule is consequential.
  • LSE pooling temperature T_ℓ = 1.5
    Fig 6 shows stability over [1.0, 2.5] and clear degradation at 0.5, so the value is safe but manually set.
  • L2 anchoring coefficient λ = 0.001
    Fixed across all experiments (A.9); no sensitivity analysis reported for this value.
  • SGD learning rate 5e-4, momentum 0.9, 3 epochs per round, 3 rounds = 5e-4, 0.9, 3, 3
    Fixed hyperparameters; no sensitivity analysis reported for LR or epoch count.
  • Patch budgets per scale (24, 6, 1) at 70x70, 224x224, 336x336 = 24 / 6 / 1
    Fig 5 shows the fine-scale count matters: removing it hurts first-round performance and 32 patches degrades results.
  • Patch scoring weights (high-frequency energy 0.4, gradient 0.3, texture 0.3) and random ratio 0.2 = 0.4 / 0.3 / 0.3, 0.2
    Hand-set in A.9; Table 12 shows limited sensitivity across strategies.
  • Label-free safety-check thresholds = not stated
    A.9 says thresholds are fixed and applied uniformly but never gives numerical values, so the rollback behavior cannot be audited.
axioms (5)
  • domain assumption Low-entropy predictions by the base detector are accurate enough to serve as pseudo-labels under unseen-generator shift
    Section 3.2 Eqs. 2-4 select reliable samples purely by prediction entropy; Table 13a shows selected pseudo-labels are about 92% accurate, but accuracy is verified only post-hoc with ground-truth labels unavailable at test time.
  • domain assumption Cross-scale consensus improves pseudo-label reliability over single-scale predictions
    Section 3.3 assumes artifacts appear at multiple resolutions and that averaging scale-level probabilities yields more reliable supervision; the ablation (Table 3, +14.1 BAcc from CSPLR) supports this empirically.
  • domain assumption Class-balanced selection prevents gradient bias without distorting the evaluation distribution
    Section 3.2 states balancing is an optimization constraint rather than a target prior estimate; the paper keeps balanced accuracy as the metric and does not reweight at evaluation.
  • ad hoc to paper The label-free safety check can detect harmful adaptation using only prediction statistics
    A.9 defines a rule on the predicted fake prior and average certainty with unspecified fixed thresholds; this is bespoke to the paper and not derived from theory.
  • standard math BCE loss with L2 anchoring toward the initial parameters prevents catastrophic drift
    Eq. 6 is standard regularized fine-tuning; the paper provides no proof that the anchoring coefficient prevents drift, only empirical stability results.

pith-pipeline@v1.3.0-alltime-deepseek · 23589 in / 24339 out tokens · 255695 ms · 2026-08-05T00:40:47.022925+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Test-Time Curriculum for Open-Set AIGC Detection." pith.science (2026). https://pith.science/paper/GGCWMA7M

@misc{pith2026260800559,
  author       = {Pith},
  title        = {Pith review of: Test-Time Curriculum for Open-Set AIGC Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGCWMA7M}},
  note         = {Machine review of arXiv:2608.00559}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

AI-generated image detectors deployed in open-world environments inevitably face distribution shifts as new and stronger generative models continue to emerge. Although existing methods improve cross-generator generalization through better representations or training data construction, they typically follow a static train-once-and-deploy paradigm and cannot adapt after deployment. In this work, we study open-set AIGC image detection from a test-time adaptation perspective. We propose Test-Time Curriculum (TTC), a simple and model-agnostic framework that adapts a detector on unlabeled test data through curriculum-based self-training. TTC starts from highly reliable pseudo-labeled samples and progressively incorporates harder yet informative cases, while enforcing class-balanced selection to reduce biased updates under generator shift. To further improve pseudo-label quality, we introduce Cross-Scale Pseudo-Label Refinement, which aggregates complementary evidence across multiple resolutions for more reliable adaptation, and applies noisy-or fusion at inference to strengthen final predictions. In addition, we construct AIGCGuard, a new benchmark containing 3,100 representative real images and 124,000 generated images from 40 of the most advanced open-source and proprietary text-to-image models. Extensive experiments on five benchmarks show that TTC substantially improves overall detection performance under diverse unseen-generator shifts, establishing a practical and effective test-time adaptation framework for open-set generated image detection.

Figures

Figures reproduced from arXiv: 2608.00559 by Jiahao Hu, Jiaxu Miao, Jingjia Mao, Jun Yu, Shuohuan Wang, Xiangzhao Hao, Yiqian Zhang, Yu Sun, Zefeng Zhang, Zhenyu Zhang, Zheyuan Gu.

Figure 1
Figure 1. Figure 1: Illustration of open-set AIGC image detection under evolving generator distributions. Existing detectors are trained [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed TTC framework. Given unlabeled test images, TTC first performs multi-scale prediction and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Robustness under varying JPEG compression, resizing, and Gaussian blur on GenImage. The experiment evaluates [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity to curriculum schedules on AIGCDetect [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity to the LSE temperature T on AIGCDe￾tect. We vary the T while keeping all other settings fixed. BAcc is reported after the first adaptation round. Patch selection strategy BAcc (%) Random 91.16 Score-only 90.07 Score + diversity 90.43 Full 90.55 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Representative real images from 16 semantic categories in AIGCGuard. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Real images and samples generated by recent text-to-image models from the same prompts. [PITH_FULL_IMAGE:figures/full_fig_p013_8.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

91 extracted references · 23 canonical work pages · 1 internal anchor

  1. [1]

    arXiv preprint arXiv:2505.14359 , year=

    Dual data alignment makes ai-generated image detector easier generalizable , author=. arXiv preprint arXiv:2505.14359 , year=

    arXiv

  2. [2]

    arXiv preprint arXiv:2509.20427 , year=

    Seedream 4.0: Toward next-generation multimodal image generation , author=. arXiv preprint arXiv:2509.20427 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2511.22699 , year=

    Z-image: An efficient image generation foundation model with single-stream diffusion transformer , author=. arXiv preprint arXiv:2511.22699 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2509.23951 , year=

    Hunyuanimage 3.0 technical report , author=. arXiv preprint arXiv:2509.23951 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2512.07584 , year=

    Longcat-image technical report , author=. arXiv preprint arXiv:2512.07584 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2602.04705 , year=

    ERNIE 5.0 Technical Report , author=. arXiv preprint arXiv:2602.04705 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2508.02324 , year=

    Qwen-image technical report , author=. arXiv preprint arXiv:2508.02324 , year=

    Pith/arXiv arXiv

  8. [8]

    Advances in neural information processing systems , volume=

    Improving robustness against common corruptions by covariate shift adaptation , author=. Advances in neural information processing systems , volume=

  9. [9]

    arXiv preprint arXiv:2006.10963 , year=

    Evaluating prediction-time batch normalization for robustness under covariate shift , author=. arXiv preprint arXiv:2006.10963 , year=

    Pith/arXiv arXiv 2006

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Test-time classifier adjustment module for model-agnostic domain generalization , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    arXiv preprint arXiv:2301.13018 , year=

    Delta: degradation-free fully test-time adaptation , author=. arXiv preprint arXiv:2301.13018 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2006.10726 , year=

    Tent: Fully test-time adaptation by entropy minimization , author=. arXiv preprint arXiv:2006.10726 , year=

    Pith/arXiv arXiv 2006

  13. [13]

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

    Efficient test-time adaptation of vision-language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  14. [14]

    International conference on machine learning , pages=

    Efficient test-time model adaptation without forgetting , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  15. [15]

    arXiv preprint arXiv:2302.12400 , year=

    Towards stable test-time adaptation in dynamic wild world , author=. arXiv preprint arXiv:2302.12400 , year=

    Pith/arXiv arXiv

  16. [16]

    International Journal of Computer Vision , volume=

    A comprehensive survey on test-time adaptation under distribution shifts , author=. International Journal of Computer Vision , volume=. 2025 , publisher=

    2025

  17. [17]

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

    Robust test-time adaptation in dynamic scenarios , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  18. [18]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Continual test-time domain adaptation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  19. [19]

    2025 IEEE International Conference on Multimedia and Expo (ICME) , pages=

    ElimPCL: Eliminating Noise Accumulation with Progressive Curriculum Labeling for Source-Free Domain Adaptation , author=. 2025 IEEE International Conference on Multimedia and Expo (ICME) , pages=. 2025 , organization=

    2025

  20. [20]

    Advances in Engineering Technology Research , volume=

    Survey for detecting AI-generated content , author=. Advances in Engineering Technology Research , volume=

  21. [21]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  22. [22]

    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=

  23. [23]

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

    Forgery-aware adaptive transformer for generalizable synthetic image detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  24. [24]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    C2p-clip: Injecting category common prompt in clip to enhance generalization in deepfake detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  25. [25]

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

    Aigi-holmes: Towards explainable and generalizable ai-generated image detection via multimodal large language models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  26. [26]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Improving synthetic image detection towards generalization: An image transformation perspective , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

    work page

  27. [27]

    Forty-first International Conference on Machine Learning , year=

    Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images , author=. Forty-first International Conference on Machine Learning , year=

    work page

  28. [28]

    arXiv preprint arXiv:2410.11835 , year=

    Aligned datasets improve detection of latent diffusion-generated images , author=. arXiv preprint arXiv:2410.11835 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2406.19435 , year=

    A sanity check for ai-generated image detection , author=. arXiv preprint arXiv:2406.19435 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2311.12397 , volume=

    Rich and poor texture contrast: A simple yet effective approach for ai-generated image detection , author=. arXiv preprint arXiv:2311.12397 , volume=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:1812.08247 , year=

    Detecting gan-generated imagery using color cues , author=. arXiv preprint arXiv:1812.08247 , year=

    Pith/arXiv arXiv

  32. [32]

    for now , author=

    CNN-generated images are surprisingly easy to spot... for now , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  33. [33]

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

    Label shift adapter for test-time adaptation under covariate and label shifts , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  34. [34]

    2019 IEEE international conference on image processing (ICIP) , pages=

    Detecting GAN-generated imagery using saturation cues , author=. 2019 IEEE international conference on image processing (ICIP) , pages=. 2019 , organization=

    work page 2019

  35. [35]

    arXiv preprint arXiv:2302.05155 , year=

    Ttn: A domain-shift aware batch normalization in test-time adaptation , author=. arXiv preprint arXiv:2302.05155 , year=

    Pith/arXiv arXiv

  36. [36]

    Test-Time Adaptation for Unsupervised Combinatorial Optimization

    Test-Time Adaptation for Unsupervised Combinatorial Optimization , author=. arXiv preprint arXiv:2601.21048 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    Architecture-Agnostic Test-Time Adaptation via Backprop-Free Embedding Alignment , author=

    work page

  38. [38]

    arXiv preprint arXiv:1903.06836 , year=

    Detecting GAN generated fake images using co-occurrence matrices , author=. arXiv preprint arXiv:1903.06836 , year=

    Pith/arXiv arXiv 1903

  39. [39]

    Advances in neural information processing systems , volume=

    Genimage: A million-scale benchmark for detecting ai-generated image , author=. Advances in neural information processing systems , volume=

  40. [40]

    Luo, Yunpeng and Du, Junlong and Yan, Ke and Ding, Shouhong , booktitle=. Lare\^

    work page

  41. [41]

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

    Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

    arXiv preprint arXiv:2402.01123 , year=

    A single simple patch is all you need for ai-generated image detection , author=. arXiv preprint arXiv:2402.01123 , year=

    Pith/arXiv arXiv

  43. [43]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    A bias-free training paradigm for more general ai-generated image detection , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  44. [44]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    SemGIR: Semantic-guided image regeneration based method for AI-generated image detection and attribution , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    work page

  45. [45]

    European Conference on Computer Vision , pages=

    Fake or jpeg? revealing common biases in generated image detection datasets , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    work page 2024

  46. [46]

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

    Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  47. [47]

    International conference on machine learning , pages=

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

    2020

  48. [48]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    A survey on curriculum learning , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2021 , publisher=

    work page 2021

  49. [49]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    The regression analysis of binary sequences , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 1958 , publisher=

    1958

  50. [50]

    Acm sigir forum , volume=

    A sequential algorithm for training text classifiers: Corrigendum and additional data , author=. Acm sigir forum , volume=. 1995 , organization=

    work page 1995

  51. [51]

    Advances in neural information processing systems , volume=

    A simple weight decay can improve generalization , author=. Advances in neural information processing systems , volume=

  52. [52]

    Advances in Neural Information Processing Systems , volume=

    Cycle self-training for domain adaptation , author=. Advances in Neural Information Processing Systems , volume=

    work page

  53. [53]

    2005 , publisher=

    Semi-supervised self-training of object detection models , author=. 2005 , publisher=

    work page 2005

  54. [54]

    Neurocomputing , volume=

    Self-training: A survey , author=. Neurocomputing , volume=. 2025 , publisher=

    2025

  55. [55]

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

    Dire for diffusion-generated image detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  56. [56]

    arXiv preprint arXiv:2504.01396 , year=

    All patches matter, more patches better: Enhance ai-generated image detection via panoptic patch learning , author=. arXiv preprint arXiv:2504.01396 , year=

    arXiv

  57. [57]

    2014 , publisher=

    Probabilistic reasoning in intelligent systems: networks of plausible inference , author=. 2014 , publisher=

    2014

  58. [58]

    Advances in neural information processing systems , volume=

    A framework for multiple-instance learning , author=. Advances in neural information processing systems , volume=

  59. [59]

    International conference on machine learning , pages=

    Attention-based deep multiple instance learning , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  60. [60]

    Electronics , volume=

    A comprehensive review on multiple instance learning , author=. Electronics , volume=. 2023 , publisher=

    work page 2023

  61. [61]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Path aggregation network for instance segmentation , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  62. [62]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Feature pyramid networks for object detection , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  63. [63]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  64. [64]

    Advances in neural information processing systems , volume=

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results , author=. Advances in neural information processing systems , volume=

  65. [65]

    International conference on machine learning , pages=

    A simple framework for contrastive learning of visual representations , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  66. [66]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Deep clustering for unsupervised learning of visual features , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  67. [67]

    arXiv preprint arXiv:1708.00489 , year=

    Active learning for convolutional neural networks: A core-set approach , author=. arXiv preprint arXiv:1708.00489 , year=

    Pith/arXiv arXiv

  68. [68]

    International conference on machine learning , pages=

    Submodularity in data subset selection and active learning , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  69. [69]

    arXiv preprint arXiv:2606.02800 , year=

    Cosmos 3: Omnimodal world models for physical ai , author=. arXiv preprint arXiv:2606.02800 , year=

    Pith/arXiv arXiv

  70. [70]

    arXiv preprint arXiv:2506.18871 , year=

    Omnigen2: Exploration to advanced multimodal generation , author=. arXiv preprint arXiv:2506.18871 , year=

    Pith/arXiv arXiv

  71. [71]

    arXiv preprint arXiv:2605.11061 , year=

    Hidream-o1-image: A natively unified image generative foundation model with pixel-level unified transformer , author=. arXiv preprint arXiv:2605.11061 , year=

    Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2505.23661 , year=

    Openuni: A simple baseline for unified multimodal understanding and generation , author=. arXiv preprint arXiv:2505.23661 , year=

    Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2511.06876 , year=

    Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions , author=. arXiv preprint arXiv:2511.06876 , year=

    arXiv

  74. [74]

    2025 , eprint=

    Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference , author=. 2025 , eprint=

    work page 2025

  75. [75]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  76. [76]

    arXiv preprint , year=

    Kolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis , author=. arXiv preprint , year=

    work page

  77. [77]

    arXiv preprint arXiv:2505.14683 , year =

    Emerging Properties in Unified Multimodal Pretraining , author =. arXiv preprint arXiv:2505.14683 , year =

    Pith/arXiv arXiv

  78. [78]

    European Conference on Computer Vision , pages=

    Pixart- : Weak-to-strong training of diffusion transformer for 4k text-to-image generation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  79. [79]

    arXiv preprint arXiv:2501.17811 , year=

    Janus-pro: Unified multimodal understanding and generation with data and model scaling , author=. arXiv preprint arXiv:2501.17811 , year=

    Pith/arXiv arXiv

  80. [80]

    arXiv preprint arXiv:2505.09568 , year=

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset , author=. arXiv preprint arXiv:2505.09568 , year=

    Pith/arXiv arXiv

Showing first 80 references.