Pith. sign in

REVIEW 1 major objections 8 minor 43 references

Pipeline discovers unknown AI face generators from rejected images

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 · glm-5.2

2026-07-09 07:18 UTC pith:FLKO4DWI

load-bearing objection Incremental unknown-generator discovery is a real contribution, but the headline clustering metrics are computed on a doubly-filtered subset and the discovery stage is carried entirely by FSD features, not I-JEPA. the 1 major comments →

arxiv 2607.07545 v1 pith:FLKO4DWI submitted 2026-07-08 cs.CV

Face-trace: Open-Set Attribution and Progressive Discovery of Synthetic Face Generators

Alessia Infantino , Claudio Schiavella , Irene Amerini This is my paper
classification cs.CV
keywords open-set recognitionsynthetic face attributionnovel category discoveryenergy-based OOD detectionforensic self-descriptorsI-JEPAHDBSCAN clusteringincremental discovery
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 proposes a three-stage pipeline for open-set synthetic face source attribution: classify images from known generators using frozen I-JEPA embeddings, reject samples from unseen generators via an energy-based out-of-distribution score, then cluster the rejected samples to discover distinct unknown generators. The central mechanism is the fusion of projected I-JEPA features with Forensic Self-Descriptors — compact descriptors that model forensic microstructures in synthetic images — to produce a representation space where unseen generators form separable clusters. The authors extend this to an incremental setting where rejected samples arrive over time: samples matching already-discovered clusters are assigned directly, while unmatched samples buffer until enough accumulate to form new clusters via HDBSCAN. On the WILD dataset of 20 synthetic face generators, the pipeline achieves 96.73% closed-set attribution accuracy, 71.25% balanced accuracy for OOD rejection, ARI=0.81 and 87.74% clustering purity for unknown-generator discovery, and 99.23% final purity in the incremental setting.

Core claim

The paper's central claim is that unknown synthetic face generators can be discovered — not merely rejected — by clustering rejected samples in a fused representation space combining projected I-JEPA features with Forensic Self-Descriptors, and that this discovery can proceed incrementally as samples arrive over time without retraining. The ablation reveals that I-JEPA embeddings alone yield ARI=0.00 for clustering unknown generators; the discovery stage's success is carried by the FSD component, which provides the forensic microstructure information needed to separate generators. The fused representation outperforms either component alone, suggesting complementarity between high-level视觉特征和法

What carries the argument

Fused representation: 64-dim PCA-reduced projected I-JEPA embedding concatenated with 64-dim PCA-reduced Forensic Self-Descriptor, forming a 128-dim vector clustered by HDBSCAN after UMAP preprocessing. Energy score from classifier logits used for OOD rejection. Incremental discovery uses Mahalanobis distance matching to existing clusters with adaptive per-cluster radii, buffering unmatched samples, and promoting candidate clusters that meet minimum support and cohesion thresholds.

Load-bearing premise

The clustering quality depends on Forensic Self-Descriptors producing compact, generator-specific clusters for unseen generators. The ablation shows that without FSD, no clustering structure emerges from I-JEPA embeddings alone. If FSD features do not generalize to future generator architectures with different forensic microstructures, the discovery stage would fail.

What would settle it

If a new class of generators produces images whose forensic microstructures are not captured by FSD — for example, architectures that leave different or weaker traces — the fused representation would not produce separable clusters, and the discovery stage would produce noise or merged groups rather than meaningful unknown-generator categories.

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

If this is right

  • A deployed forensic system could progressively build a catalog of unknown generators without retraining, assigning pseudo-labels to new sources as they appear in a stream.
  • The incremental design means the system can operate in streaming conditions, processing rejected samples one at a time rather than requiring a batch of unknowns upfront.
  • The modular separation of classification, rejection, and discovery allows each component to be upgraded independently — a better rejection module would directly improve discovery input quality.
  • The cross-dataset experiment (two novel clusters promoted, ARI=0.33, purity=78.94%) suggests the discovered unknown space can transfer across data distributions, though with significant degradation.

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

1 major / 8 minor

Summary. The paper proposes a pipeline for open-set synthetic face source attribution combining three stages: (1) closed-set classification using frozen I-JEPA embeddings, (2) energy-based OOD rejection, and (3) unknown generator discovery via HDBSCAN clustering of rejected samples using fused I-JEPA + Forensic Self-Descriptor (FSD) features. The pipeline is extended to an incremental setting where rejected samples arrive over time and are either matched to existing discovered clusters or buffered for new cluster promotion. Experiments on the WILD dataset (10 known, 10 unknown generators) report 96.73% closed-set accuracy, 71.25% balanced rejection accuracy, ARI=0.81/NMI=0.90/87.74% purity for offline clustering, and 99.23% final purity in the incremental setting. Cross-dataset and post-processing robustness experiments are also reported.

Significance. The paper addresses a practically important problem: organizing rejected unknown-generator samples into coherent groups rather than simply discarding them. The incremental, non-transductive formulation is a meaningful contribution to the forensic attribution literature. The ablation in Table VII is commendably honest in showing that I-JEPA embeddings alone yield ARI=0.00 and that the discovery stage's clustering quality is carried by FSD features. The cross-dataset experiment (Table X, Appendix C) and post-processing robustness analysis (Table VI) provide useful negative results. The pipeline is well-motivated and the experimental protocol is generally thorough.

major comments (1)
  1. §V-B and §VII-C: The headline clustering metrics (ARI=0.81, NMI=0.90, purity=87.74%) are computed only on non-noise samples, as stated in §V-B: 'all clustering metrics are computed only on non-noise samples.' The paper does not report what fraction of the 10,000 OOD samples are marked as noise by HDBSCAN. Additionally, the energy rejection stage (Table III) achieves TPR=47.97%, meaning approximately 52% of OOD samples are incorrectly accepted as known and never reach the discovery stage. The headline clustering metrics are therefore computed on a doubly-filtered subset of unknown samples. This is load-bearing for the central claim of 'discovering unknown generators.' The authors should report (a) the noise fraction, (b) the effective number of samples on which clustering metrics are computed, and (c) clustering metrics computed on the full OOD population (with noise samples counted as a
minor comments (8)
  1. Table III: The TPR of 47.97% means a majority of OOD samples bypass the discovery stage entirely. This should be stated explicitly in the main text as a known limitation of the pipeline, not only inferable from the table.
  2. §IV-D: The Mahalanobis distance is used for matching, but it is unclear whether the covariance matrix is estimated per-cluster or globally. Please clarify.
  3. §VI-F: The matching radius is defined using the 0.95 quantile of cluster distances and a margin equal to 0.95. It is unclear how these two values combine — is the radius the quantile value, the quantile plus the margin, or the quantile multiplied by the margin? Please clarify.
  4. Table VI vs. Table VII: The 'Plain' row in Table VI reports ARI=0.87, NMI=0.94, purity=89.19%, while Table VII's default row reports ARI=0.79, NMI=0.88, purity=0.86 for the same configuration (all open-set samples, no rejection, same HDBSCAN parameters). Please reconcile these differences.
  5. §VII-E: The in-the-wild experiment reports ARI=0.33, NMI=0.40, purity=78.94% for promoted novel clusters (Appendix XI), which is substantially lower than the WILD results. The main text should discuss this performance gap more explicitly.
  6. The paper uses both 'Forensic Self-Descriptors' and 'Forensic Self-Descriptions' (§II-E). Please use consistent terminology.
  7. Fig. 4 and Fig. 5: The UMAP visualizations would benefit from explicit legend labels for all generator classes.
  8. §III-C, Eq. (8): The per-cluster matching radius δ_q is mentioned but its computation is only described later in §VI-F. A forward reference would improve clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

The referee raises a valid and important point about the transparency of the headline clustering metrics. We agree that the noise fraction and effective sample counts must be reported, and that full-population clustering metrics should be included. We will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: §V-B and §VII-C: The headline clustering metrics (ARI=0.81, NMI=0.90, purity=87.74%) are computed only on non-noise samples, as stated in §V-B: 'all clustering metrics are computed only on non-noise samples.' The paper does not report what fraction of the 10,000 OOD samples are marked as noise by HDBSCAN. Additionally, the energy rejection stage (Table III) achieves TPR=47.97%, meaning approximately 52% of OOD samples are incorrectly accepted as known and never reach the discovery stage. The headline clustering metrics are therefore computed on a doubly-filtered subset of unknown samples. This is load-bearing for the central claim of 'discovering unknown generators.' The authors should report (a) the noise fraction, (b) the effective number of samples on which clustering metrics are computed, and (c) clustering metrics computed on the full OOD population (with noise samples counted as a

    Authors: The referee is correct on all counts. The current presentation is incomplete in a way that could mislead readers about the effective coverage of the discovery stage. We will address this in the revision as follows: revision_made = 'yes'. revision_details = (1) We will report the noise fraction from HDBSCAN on the energy-rejected samples. (2) We will report the effective number of samples on which the headline clustering metrics (ARI, NMI, purity) are computed, making explicit the chain: 10,000 OOD samples → ~4,797 correctly rejected by energy (TPR=47.97%) → subset marked as non-noise by HDBSCAN → metrics computed on this subset. (3) We will add clustering metrics computed on the full OOD population of 10,000 samples, treating both (i) OOD samples incorrectly accepted as known by the energy stage and (ii) HDBSCAN noise samples as assignment failures (i.e., each counted as belonging to a single 'unassigned' group or as singleton clusters, following standard practice for evaluating clustering with noise). This will give readers a complete picture of end-to-end discovery performance, including the compounding effect of rejection errors. We agree that the abstract and main text should contextualize the headline numbers with these effective coverage figures so that the central claim of 'discovering unknown generators' is not overstated. We note that the incremental setting results (Table V) already report assignment rates (DMR, FAR, SBR) that partially address coverage, but the offline clustering section does not, and this gap will be closed. revision: no

Circularity Check

0 steps flagged

No significant circularity found; the derivation chain is self-contained against external benchmarks.

full rationale

The paper's pipeline combines four components: (1) a frozen I-JEPA encoder [30] for feature extraction, (2) an MLP classifier trained with cross-entropy on known generators, (3) energy-based OOD rejection [12] with a threshold calibrated on known-generator validation data at 5% FPR, and (4) HDBSCAN clustering [31] on fused FSD+I-JEPA features for unknown generator discovery. Each component is either a standard method from external work or a straightforward supervised training step. The key representation for discovery—Forensic Self-Descriptors [28]—is cited from Nguyen, Azizpour, and Stamm, who are not authors of the present paper (Infantino, Schiavella, Amerini). Similarly, the autonomous clustering system [29] is by the same external group. The ablation in Table VII transparently shows that I-JEPA features alone yield ARI=0.00 and that FSD carries the clustering performance, but this is a design-dependency and generalization concern, not circularity: the FSD features are externally developed and the clustering targets (unknown generators) are held out during training. The skeptic's concern about metrics being computed only on non-noise samples after OOD rejection is a valid evaluation-validity issue (potential overstatement of practical performance), but it does not constitute circular derivation—no step reduces to its own inputs by construction, no prediction is a fitted parameter renamed, and no load-bearing self-citation chain exists. The score of 1 reflects the minor observation that the central discovery result depends heavily on an externally cited feature extractor whose generalization is untested beyond the WILD dataset, but this is normal scientific dependency, not circularity.

Axiom & Free-Parameter Ledger

15 free parameters · 5 axioms · 0 invented entities

The paper introduces no new theoretical entities, particles, or forces. It combines existing components (I-JEPA, FSD, energy scores, HDBSCAN, UMAP) into a pipeline. The free parameters are numerous (15+), mostly clustering and buffer hyperparameters, with limited sensitivity analysis beyond HDBSCAN parameters (Table IX) and generator splits (Table VIII). The promotion threshold (ρ_max=26) and buffer size (B=700) appear to be tuned to the WILD dataset without independent justification.

free parameters (15)
  • Energy temperature T = 1
    Set to 1 in §VI-D without tuning; standard default from the energy OOD method.
  • Rejection threshold τ = 5% FPR on validation set
    Selected on validation set by fixing false positive rate at 5% (§IV-C). This is a standard calibration choice but directly controls the ID/OOD tradeoff.
  • UMAP n_neighbors = 15 (offline), 20 (init)
    Chosen for clustering pre-processing (§VI-E); no justification or sensitivity analysis provided.
  • UMAP min_dist = 0.3
    Chosen for clustering pre-processing (§VI-E); no justification provided.
  • UMAP n_components = 32
    Output dimensionality for clustering pre-processing (§VI-E); no justification provided.
  • HDBSCAN min_cluster_size = 90 (offline), 40 (buffer)
    Clustering parameter (§VI-E, §VI-F); ablation in Table IX shows sensitivity.
  • HDBSCAN min_samples = 15 (offline), 5 (buffer)
    Clustering parameter (§VI-E, §VI-F); ablation in Table IX shows sensitivity.
  • HDBSCAN epsilon (cluster_selection_epsilon) = 0.5 (offline), 0.6 (init/buffer)
    Clustering parameter (§VI-E, §VI-F); ablation in Table IX shows sensitivity.
  • Buffer size B = 700
    Buffer processing threshold (§VI-F); no sensitivity analysis provided.
  • Minimum promotion size n_min = 300 (WILD), 200 (in-the-wild)
    Minimum samples for cluster promotion (§VI-F, §VI-G); reduced for external data without justification beyond fewer samples.
  • Maximum promotion dispersion ρ_max = 26
    Cohesion threshold for cluster promotion (§VI-F); no sensitivity analysis or derivation provided.
  • Matching radius quantile = 0.95
    Per-cluster Mahalanobis distance quantile for matching (§VI-F); margin also set to 0.95.
  • Max buffer retention = 4 buffer clustering attempts
    Samples removed after 4 attempts (§VI-F); no sensitivity analysis provided.
  • PCA dimensions per branch = 64
    Both I-JEPA and FSD reduced to 64 dims before concatenation (§IV-D); no justification provided.
  • Label smoothing = 0.1
    Cross-entropy label smoothing (§VI-B); standard value, no tuning mentioned.
axioms (5)
  • domain assumption Frozen I-JEPA embeddings preserve generator-specific information distinguishable by a lightweight classifier
    §IV-A motivates I-JEPA choice by arguing that synthetic faces share semantic structure, so high-level features may preserve source differences. The ablation (Table VII) partially contradicts this: raw I-JEPA yields ARI=0.00 for clustering.
  • domain assumption Energy scores from a classifier trained on known generators separate ID from OOD samples
    §IV-C assumes known generators produce lower energy. The 71.25% balanced accuracy and post-processing degradation (Table VI) show this separation is imperfect.
  • domain assumption FSD features capture forensic microstructures that cluster by generator
    §IV-D combines I-JEPA with FSD based on this assumption. Table VII shows FSD alone achieves ARI=0.66, supporting it, but generalization to unseen architectures is untested.
  • standard math HDBSCAN can discover generator-specific clusters without knowing the number of generators
    HDBSCAN is a standard density-based clustering method that does not require k. The paper uses it appropriately (§IV-D).
  • domain assumption Rejected samples from the same unknown generator form compact clusters in the fused embedding space
    §III-C defines discovery as partitioning rejected samples into clusters. The Mahalanobis matching and promotion criteria (§IV-D, §IV-E) assume cluster compactness.

pith-pipeline@v1.1.0-glm · 21869 in / 3708 out tokens · 574532 ms · 2026-07-09T07:18:24.258697+00:00 · methodology

0 comments
read the original abstract

Recent advances in generative Artificial Intelligence have made synthetic face images increasingly realistic, creating new challenges for multimedia forensics. Source attribution methods should not only identify the generator of an image when the source is known, but also handle samples produced by previously unseen models. However, most existing approaches address synthetic face attribution in a closed-set setting, where all possible generators are available during training. This assumption does not hold in real-world scenarios, where new generators continuously appear and rejected samples should be organized rather than simply discarded. In this work we propose a pipeline for open-set synthetic face source attribution that combines known generator classification, energy-based OOD rejection, and unknown generator discovery. A classifier is trained on known generators using frozen I-JEPA embeddings, while rejected samples are represented by combining projected I-JEPA features with Forensic Self-Descriptors and then clustered to discover groups of unknown generators. We also extend the discovery stage to an incremental scenario, where rejected samples arrive over time. Experiments on the WILD dataset show that the proposed method achieves 96.73% closed-set attribution accuracy. In the open-set setting, energy-based rejection reaches 71.25% balanced accuracy, while rejected samples are clustered into meaningful unknown-generator groups, obtaining an ARI of 0.81, an NMI of 0.90, and an overall clustering purity of 87.74%. In the incremental setting, the discovered generator space is progressively extended while maintaining a final purity of 99.23%. Cross-dataset experiments suggest that the pipeline can operate beyond the original dataset distribution, although post-processing remains challenging.

Figures

Figures reproduced from arXiv: 2607.07545 by Alessia Infantino, Claudio Schiavella, Irene Amerini.

Figure 1
Figure 1. Figure 1: Overview of the proposed offline pipeline. Given an input image, the frozen I-JEPA target encoder extracts a raw embedding, which is processed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Incremental discovery logic. E. Pseudo-Label Assignment Once a candidate cluster is promoted, it receives a pseudo￾label and becomes part of the reliable unknown space used for future matching. Future rejected samples will be compared against this set during the matching stage. If a new rejected sample is com￾patible with a promoted cluster, it receives the corresponding pseudo-label instead of being inser… view at source ↗
Figure 3
Figure 3. Figure 3: Closed-Set Attribution Confusion Matrix. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP visualization of the clustering results on energy-rejected samples. (a) HDBSCAN cluster assignments (labelled by dominant generator). (b) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: UMAP visualization of the reliable unknown space before and after clusters promotion. Points are colored using ground-truth generator labels. (a) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Binary matrix of accepted and rejected ID and in-the-wild OOD [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Contingency matrix of the assigned samples in the final reliable [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: UMAP visualization of the reliable unknown space before and after cluster promotion. (a) Initial reliable unknown space, containing clusters discovered [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: UMAP visualization of the initialized space in the incremental experiment. (a) HDBSCAN cluster assignments. (b) Ground-truth generator labels. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Evolution of the buffer during incremental discovery across 3 discovery attempts. The top row shows the buffered samples colored according to the [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Final contingency matrix in the in-the-wild incremental setting. Most wild known ID samples are correctly assigned to their corresponding ID [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ground Truth UMAP visualization of the initial and final space in the in the wild incremental experiment. (a) Initial space. (b) Final space with [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Binary matrices of accepted and rejected ID and OOD post-processed samples using the energy rejection score. (a) 1-step post-processing. (b) 2-step [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Energy distribution and box plot of post-processed closed-set samples vs post-processed open-set samples. (a) 1-step post-processing. (b) 2-step [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Energy distribution and box plot of original closed-samples vs post-processed closed-set samples. (a) 1-step post-processing. (b) 2-step post-processing. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: UMAP visualizations of WILD open-set generator space using different embedding representations. Points are colored according to the ground-truth [PITH_FULL_IMAGE:figures/full_fig_p017_16.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

43 extracted references · 43 canonical work pages · 13 internal anchors

  1. [1]

    AI-synthesized faces are indistin- guishable from real faces and more trustworthy,

    S. J. Nightingale and H. Farid, “AI-synthesized faces are indistin- guishable from real faces and more trustworthy,”Proc. Natl. Acad. Sci. U.S.A., vol. 119, no. 8, Art. no. e2120481119, Feb. 2022, doi: 10.1073/pnas.2120481119

    work page doi:10.1073/pnas.2120481119 2022

  2. [2]

    Face Deepfakes -- A Comprehensive Review

    T. Fernando, D. Priyasad, S. Sridharan, A. Ross, and C. Fookes, “Face deepfakes: A comprehensive review,”arXiv preprint arXiv:2502.09812, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Deepfake Media Forensics: State of the Art and Challenges Ahead

    I. Amerini, M. Barni, S. Battiato, P. Bestagini, G. Boato, T. S. Bonaven- tura, V . Bruni, R. Caldelli, F. De Natale, R. De Nicola, L. Guarnera, S. Mandelli, G. L. Marcialis, M. Micheletto, A. Montibeller, G. Orr `u, A. Ortis, P. Perazzo, G. Puglisi, D. Salvi, S. Tubaro, C. M. Tonti, M. Villari, and D. Vitulano, “Deepfake media forensics: State of the art...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Deepfake attribution: On the source identification of artificially generated images,

    B. Khoo, R. C.-W. Phan, and C.-H. Lim, “Deepfake attribution: On the source identification of artificially generated images,”WIREs Data Mining Knowl. Discov., 2021, doi: 10.1002/widm.1438

    work page doi:10.1002/widm.1438 2021

  5. [5]

    Deepfake detection, attribution, and authentication: Insights from the FF4ALL project,

    I. Amerini, M. Barni, S. Battiato, P. Bestagini, G. Boato, P. Bongini, V . Bruni, R. Casula, L. Cirillo, R. Caldelli, G. Daidone, F. De Natale, R. De Nicola, L. Guarnera, S. M. La Cava, S. Mandelli, G. L. Marcialis, M. Micheletto, A. Montibeller, V . Negroni, G. Orr`u, P. Perazzo, G. Puglisi, D. Salvi, B. Tondi, S. Tubaro, M. Villari, and D. Vitulano, “De...

    work page 2026

  6. [6]

    A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges

    M. Salehi, H. Mirzaei, D. Hendrycks, Y . Li, M. H. Rohban, and M. Sabokrou, “A unified survey on anomaly, novelty, open-set, and out-of- distribution detection: Solutions and future challenges,”arXiv preprint arXiv:2110.14051, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  7. [7]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    L. McInnes, J. Healy, and J. Melville, “UMAP: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  8. [8]

    Generalized out-of-distribution detection: A survey,

    J. Yang, K. Zhou, Y . Li, and Z. Liu, “Generalized out-of-distribution detection: A survey,”Int. J. Comput. Vis., vol. 132, pp. 5635–5662, 2024, doi: 10.1007/s11263-024-02117-4

    work page doi:10.1007/s11263-024-02117-4 2024

  9. [9]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks,

    D. Hendrycks and K. Gimpel, “A baseline for detecting misclassified and out-of-distribution examples in neural networks,” inProceedings of the 5th International Conference on Learning Representations (ICLR), 2017

    work page 2017

  10. [10]

    Scaling Out-of-Distribution Detection for Real-World Settings

    D. Hendrycks, S. Basart, M. Mazeika, M. Mostajabi, J. Steinhardt, and D. Song, “A benchmark for anomaly segmentation,”arXiv preprint arXiv:1911.11132, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  11. [11]

    On the Importance of Gradients for Detecting Distributional Shifts in the Wild

    R. Huang, A. Geng, and Y . Li, “On the importance of gradi- ents for detecting distributional shifts in the wild,”arXiv preprint arXiv:2110.00218, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  12. [12]

    Energy-based out-of- distribution detection,

    W. Liu, X. Wang, J. D. Owens, and Y . Li, “Energy-based out-of- distribution detection,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 21464–21475

    work page 2020

  13. [13]

    GEN: Pushing the limits of softmax- based out-of-distribution detection,

    X. Liu, Y . Lochman, and C. Zach, “GEN: Pushing the limits of softmax- based out-of-distribution detection,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 23946–23955, 2023

    work page 2023

  14. [14]

    Towards generalizable detector for generated image,

    Q. Cai, C. Wu, Y . Zhang, J. Yu, and X. Tian, “Towards generalizable detector for generated image,”OpenReview, 2026. [Online]. Available: https://openreview.net/forum?id=MeawZGFIcT

    work page 2026

  15. [15]

    Deepfake network architecture attribution,

    T. Yang, Z. Huang, J. Cao, L. Li, and X. Li, “Deepfake network architecture attribution,” inProc. AAAI Conf. Artif. Intell., vol. 36, no. 4, pp. 4662–4670, 2022, doi: 10.1609/aaai.v36i4.20391

    work page doi:10.1609/aaai.v36i4.20391 2022

  16. [16]

    RepMix: Representation mixing for robust attribution of synthesized images,

    T. Bui, N. Yu, and J. Collomosse, “RepMix: Representation mixing for robust attribution of synthesized images,” inProceedings of the European Conference on Computer Vision (ECCV), 2022, pp. 146–163

    work page 2022

  17. [17]

    Progres- sive open space expansion for open-set model attribution,

    T. Yang, D. Wang, F. Tang, X. Zhao, J. Cao, and S. Tang, “Progres- sive open space expansion for open-set model attribution,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 15856– 15865, 2023

    work page 2023

  18. [18]

    Beyond deepfake vs real: Facial deepfake detection in the open-set paradigm,

    T. T. N. Bahavan, S. Seneviratne, S. Saha, K. Chen, S. Rasnayaka, and S. Halgamuge, “Beyond deepfake vs real: Facial deepfake detection in the open-set paradigm,”arXiv preprint arXiv:2503.08055, 2026

    work page arXiv 2026

  19. [19]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProceedings of the 38th International Conference on Machine Learning (ICML), 2021, vol. 139, pp. 8748–8763

    work page 2021

  20. [20]

    DINOv3

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. J ´egou, P. Labatut, and P. Bojanowski, “DINOv3,”arXiv preprint arXiv:2508.10104, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Supervised contrastive learning for few-shot AI-generated image detection and attribution,

    J. A. Urue ˜na, D. Camacho, and J. Huertas Tato, “Supervised contrastive learning for few-shot AI-generated image detection and attribution,” arXiv preprint arXiv:2511.16541, 2025

    work page arXiv 2025

  22. [22]

    Generalizable synthetic image detection via language-guided contrastive learning,

    H. Wu, J. Zhou, and S. Zhang, “Generalizable synthetic image detection via language-guided contrastive learning,”IEEE Trans. Artif. Intell., vol. 7, no. 6, pp. 3485–3496, 2026, doi: 10.1109/TAI.2025.3641104

    work page doi:10.1109/tai.2025.3641104 2026

  23. [23]

    DE-FAKE: Detection and attribution of fake images generated by text-to-image generation models,

    Z. Sha, Z. Li, N. Yu, and Y . Zhang, “DE-FAKE: Detection and attribution of fake images generated by text-to-image generation models,” inProc. ACM SIGSAC Conf. Comput. Commun. Secur. (CCS), pp. 3418–3432, 2023, doi: 10.1145/3576915.3616588

    work page doi:10.1145/3576915.3616588 2023

  24. [24]

    Are CLIP features all you need for universal synthetic image origin attribution?,

    D. Cioni, C. Tzelepis, L. Seidenari, and I. Patras, “Are CLIP features all you need for universal synthetic image origin attribution?,” inCom- puter Vision–ECCV 2024 Workshops, pp. 363–382, Cham, Switzerland: Springer, 2025, doi: 10.1007/978-3-031-92648-8 22

    work page doi:10.1007/978-3-031-92648-8 2024

  25. [25]

    Zero-shot clustering of embeddings with pretrained and self-supervised learning encoders,

    S. C. Lowe, J. B. Haurum, S. Oore, T. B. Moeslund, and G. W. Taylor, “Zero-shot clustering of embeddings with pretrained and self-supervised learning encoders,”OpenReview, 2024

    work page 2024

  26. [26]

    Is it all a cluster game? -- Exploring Out-of-Distribution Detection based on Clustering in the Embedding Space

    P. Sinhamahapatra, R. Koner, K. Roscher, and S. G ¨unnemann, “Is it all a cluster game? Exploring out-of-distribution detection based on clustering in the embedding space,”arXiv preprint arXiv:2203.08549, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  27. [27]

    Towards discovery and attribution of open-world GAN generated images,

    S. Girish, S. Suri, S. S. Rambhatla, and A. Shrivastava, “Towards discovery and attribution of open-world GAN generated images,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 14094–14103, 2021

    work page 2021

  28. [28]

    Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Clustering of AI-generated Images

    T. D. Nguyen, A. Azizpour, and M. C. Stamm, “Forensic self- descriptions are all you need for zero-shot detection, open-set source attribution, and clustering of AI-generated images,”arXiv preprint arXiv:2503.21003, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    Autonomous and Self-Adapting System for Synthetic Media Detection and Attribution

    A. Azizpour, T. D. Nguyen, and M. C. Stamm, “Autonomous and self- adapting system for synthetic media detection and attribution,”arXiv preprint arXiv:2504.03615, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Self-supervised learning from images with a joint-embedding predictive architecture,

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rab- bat, Y . LeCun, and N. Ballas, “Self-supervised learning from images with a joint-embedding predictive architecture,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 15619–15629

    work page 2023

  31. [31]

    Density-based clustering based on hierarchical density estimates,

    R. J. G. B. Campello, D. Moulavi, and J. Sander, “Density-based clustering based on hierarchical density estimates,” inAdvances in Knowledge Discovery and Data Mining, J. Pei, V . S. Tseng, L. Cao, H. Motoda, and G. Xu, Eds. Berlin, Germany: Springer, 2013, vol. 7819, pp. 160–172, doi: 10.1007/978-3-642-37456-2 14

    work page doi:10.1007/978-3-642-37456-2 2013

  32. [32]

    WILD: a new in-the-Wild Image Linkage Dataset for synthetic image attribution

    P. Bonginiet al., “WILD: A new in-the-wild image linkage dataset for synthetic image attribution,”arXiv preprint arXiv:2504.19595, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    SFHQ-T2I: Synthetic faces from text-to-image models,

    D. Beniaguev, “SFHQ-T2I: Synthetic faces from text-to-image models,” Kaggle, 2024, doi: 10.34740/KAGGLE/DSV/9548853

    work page doi:10.34740/kaggle/dsv/9548853 2024

  34. [34]

    So-Fake: Benchmarking and explaining social media image forgery detection,

    Z. Huang, T. Li, X. Li, H. Wen, Y . He, J. Zhang, H. Fei, X. Yang, X. Huang, B. Peng, and G. Cheng, “So-Fake: Benchmarking and explaining social media image forgery detection,”arXiv preprint arXiv:2505.18660, 2025

    work page arXiv 2025

  35. [35]

    AI-Face: A Million-Scale Demographically Annotated AI-Generated Face Dataset and Fairness Benchmark

    L. Lin, Santosh, M. Wu, X. Wang, and S. Hu, “AI-Face: A million- scale demographically annotated AI-generated face dataset and fairness benchmark,”arXiv preprint arXiv:2406.00783, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    Supervised contrastive learning,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” inAdvances in Neural Information Processing Systems, vol. 33, pp. 18661–18673, 2020

    work page 2020

  37. [37]

    Learning to discover novel visual categories via deep transfer clustering,

    K. Han, A. Vedaldi, and A. Zisserman, “Learning to discover novel visual categories via deep transfer clustering,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 8400–8408

    work page 2019

  38. [38]

    Generalized category discovery,

    S. Vaze, K. Han, A. Vedaldi, and A. Zisserman, “Generalized category discovery,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  39. [39]

    Parametric classification for generalized category discovery: A baseline study,

    X. Wen, B. Zhao, and X. Qi, “Parametric classification for generalized category discovery: A baseline study,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 16544– 16554

    work page 2023

  40. [40]

    ProtoGCD: Unified and unbiased prototype learning for generalized category discovery,

    S. Ma, F. Zhu, X.-Y . Zhang, and C.-L. Liu, “ProtoGCD: Unified and unbiased prototype learning for generalized category discovery,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 7, pp. 6022–6038, 2025, doi: 10.1109/TPAMI.2025.3557502

    work page doi:10.1109/tpami.2025.3557502 2025

  41. [41]

    Open-world deepfake attribution via confidence-aware asymmetric learning,

    H. Zheng, N. Pu, W. Li, T. Long, N. Sebe, and Z. Zhong, “Open-world deepfake attribution via confidence-aware asymmetric learning,”arXiv preprint arXiv:2512.12667, 2025

    work page arXiv 2025

  42. [42]

    VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking

    R. Du, D. Chang, K. Liang, T. Hospedales, Y .-Z. Song, and Z. Ma, “On-the-fly category discovery,” inProc. IEEE/CVF Conf. Com- put. Vis. Pattern Recognit. (CVPR), pp. 11691–11700, 2023, doi: 10.1109/CVPR52729.2023.01125

    work page doi:10.1109/cvpr52729.2023.01125 2023

  43. [43]

    Novel Class Discovery: an Introduction and Key Concepts

    C. Troisemaine, V . Lemaire, S. Gosselin, A. Reiffers-Masson, J. Flocon- Cholet, and S. Vaton, “Novel class discovery: An introduction and key concepts,”arXiv preprint arXiv:2302.12028, 2023. PREPRINT. UNDER REVIEW. 13 APPENDIXA CLUSTERING OF UNKNOWN GENERATORS This appendix reports additional results supporting the clus- tering analysis discussed in the ...

    work page internal anchor Pith review Pith/arXiv arXiv 2023