Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Deepfake detection works better when the model plans which face parts to inspect, then examines each with independent forensic evidence rather than inventing its own explanations.

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 · grok-4.5

2026-07-13 20:46 UTC pith:SIEVRWWR

load-bearing objection Solid engineering of stage-gated part evidence plus a useful 5-level benchmark; the independence claim is softer than advertised because the same forensic encoders help write the SFT targets. the 3 major comments →

arxiv 2603.21526 v2 pith:SIEVRWWR submitted 2026-03-23 cs.CV

VIGIL: Part-Grounded Structured Reasoning for Generalizable Deepfake Detection

Xinghan Li , Junhao Xu , Jingjing Chen This is my paper
classification cs.CV
keywords deepfake detectionmultimodal large language modelspart-grounded reasoningforensic signal injectiongeneralization benchmarkreinforcement learningface parsing
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.

Current multimodal language models that try to explain deepfakes blur observation and invention: they generate both the claim and the supporting “evidence” in one pass, so it is hard to tell real forensic analysis from fluent hallucination. VIGIL restructures that process into a plan-then-examine pipeline inspired by human forensics. The model first decides, from a global look, which facial parts deserve scrutiny; only afterward does it receive independently extracted frequency-domain and pixel-level signals for those parts, via a stage-gated injection that keeps planning free of external bias. A three-stage training regime—supervised annotation, hard-sample self-training, and reinforcement learning with part-aware and evidence-conclusion rewards—forces anatomical validity and logical coherence. The authors also release OmniFake, a five-level benchmark that trains detectors on only three basic generators and then tests them all the way to in-the-wild social-media images. Across every level, and on an external benchmark, VIGIL outperforms both specialized binary detectors and other explanation-producing models while occasionally reversing an initially wrong global judgment once part-level evidence arrives.

Core claim

Decoupling claim generation from independently sourced, region-specific forensic evidence—via a plan-then-examine pipeline with stage-gated part-level signal injection and part-aware reinforcement learning—yields more accurate and more generalizable deepfake detection than either black-box expert detectors or prior multimodal language-model methods that entangle evidence and conclusion.

What carries the argument

The plan-then-examine pipeline with context-aware dynamic signal injection: eight fixed anatomical part tokens receive mask-pooled frequency and pixel embeddings only inside the examination stage, never during planning, so part selection remains the model’s own perceptual decision.

Load-bearing premise

That a fixed face parser plus two external forensic encoders can supply truly independent, still-diagnostic region evidence even for generators and real-world degradations never seen in training.

What would settle it

Train VIGIL only on the three foundational generators, then measure accuracy and explanation coherence on a new generator family or heavy-compression regime where the face parser fails or the spectral/pixel branches become uninformative; if the gain over non-part-centric baselines disappears, the decoupling claim fails.

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

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

3 major / 6 minor

Summary. VIGIL proposes a plan-then-examine MLLM pipeline for interpretable deepfake detection: the model first selects facial parts from global cues, then examines each part with stage-gated injection of independently extracted spectral and pixel-level forensic embeddings (Eqs. 1–3). Training proceeds in three stages—signal-semantic SFT, hard-sample rejection sampling, and GRPO with part-aware and evidence–conclusion consistency rewards. The authors also introduce OmniFake, a 5-level hierarchical benchmark on which a model trained only on StyleGAN2, SD 1.4, and SimSwap is evaluated up to in-the-wild social-media data. Reported results show 93.1% average accuracy on OmniFake and 92.1% on HydraFake, exceeding retrained expert detectors and concurrent MLLM detectors, with ablations attributing gains to part-centric structure, stage-gating, dual forensic branches, and the forensic-specific rewards.

Significance. If the claims hold, the work advances both evaluation practice and method design for generalizable, explainable deepfake detection. OmniFake’s minimal-prior, progressive protocol is a concrete improvement over the common FF++→few-benchmark setup and is independently useful. The plan-then-examine architecture with stage-gated part evidence, the qualitative reasoning-reversion behavior (Fig. 4), and the multi-dimensional forensic rewards are substantive engineering contributions relative to post-hoc explanation MLLMs. Strengths include thorough retraining of expert baselines, reconstruction of concurrent MLLM training data on the same split, component ablations (Tables 3–6), robustness under compression/blur (Table 8), and external HydraFake evaluation. The main open question is whether the claimed decoupling of claims from independently sourced evidence is fully supported by the training pipeline, or whether performance largely reflects multi-signal fusion plus structured output.

major comments (3)
  1. [Sec. 4.1–4.2, Fig. 3, Eqs. 1–3] Sec. 4.1–4.2 and Fig. 3: The central claim that stage-gated injection supplies evidence independent of planning is undercut by the Stage-1 annotation pipeline. Multi-signal extraction ranks Top-K ROIs with the same spectral/pixel encoders that later produce ek; the expert LLM then synthesizes those scores into the five-part SFT targets. Supervised trajectories therefore already encode which parts the forensic branches flag as anomalous. At inference, injecting the identical ek only inside <part_evidence> may largely retrieve a memorized signal-to-structure mapping rather than perform mid-chain examination of independent evidence. Table 3’s complementarity (+8.4% combined) and Fig. 4’s reversion are then ambiguous. Please either (i) retrain SFT annotations without forensic ROI ranking (visual descriptions only, or random/shuffled part targets) and re-report L1–L5, or (ii) substantially qu
  2. [Sec. 4.2, Eq. (4); Tables 6–7] Eq. (4), Sec. 4.2: R_cons grants reward only when an independent judge LLM recovers the correct label from evidence text excluding the answer. Because Stage-1/2 trajectories were written to be label-coherent by construction, the evidence language can carry an implicit verdict (e.g., “periodic anomalies inconsistent with organic skin”), so the judge may be reading stylistic cues rather than verifying forensic grounding. This weakens the claim that R_cons enforces evidence–conclusion coherence over template memorization. Please report (a) judge accuracy on evidence from the Cold-Start model vs. full model, (b) a control where evidence is paraphrased or stripped of evaluative language, and/or (c) inter-judge agreement; otherwise soften the interpretation of R_cons in Sec. 5.3–5.4.
  3. [Sec. 4.1; Table 1 (DFDC); Table 8] Sec. 4.1 (K=8 masks, face parsing [75]) and L5 results: Spatial existence checks in R_part and mask-guided pooling (Eq. 1) assume reliable anatomical masks. Under L5 degradations (compression, resizing, occlusion in DFDC/Chameleon/WildRF), parser failures would both corrupt ek and mis-penalize valid plans. Table 1 shows a large drop on DFDC (72.2%) relative to other L5 sets. Please quantify parser failure rates or mask IoU under the L5 corruptions used in Table 8, and report accuracy conditioned on successful vs. failed parsing; if failure is non-negligible, discuss how the “anatomical validity” reward and part-level evidence remain meaningful.
minor comments (6)
  1. [Tables 1, 3–6] Table 1 vs. Tables 3–6: Avg in Table 1 is a per-dataset micro-style average, while Avg† is the mean of five level-wise averages. State this explicitly in the Table 1 caption and when comparing Cold-Start numbers across tables to avoid apparent inconsistencies (e.g., 89.8 vs. 90.6).
  2. [Sec. 4.1, Eq. (3)] Eq. (3): α and γ are zero-initialized learnable scalars; report their learned magnitudes (or norms of injected residuals) after each stage so readers can judge how strongly forensic signals actually enter the token space.
  3. [Sec. 5.1; Tables 1–2] Sec. 5.1: “Veritas (ICLR’26)” and similar future-dated venue tags will confuse readers; use arXiv IDs or “under review / concurrent” until camera-ready.
  4. [Fig. 2, Fig. 3] Fig. 2 / Fig. 3: “Partern Injection” and “Global Aanalysis” appear to be typos; also standardize <part_evidence> vs. “Examine” terminology between text and figures.
  5. [Sec. 3.1–3.2] Sec. 3.1: Exact train/test identity de-duplication threshold and per-level image counts beyond the approximate 105K/105K and level totals would aid reproducibility; point to the supplement more precisely.
  6. [Sec. 2.2] Related Work: briefly position against concurrent region-aligned MLLM detectors (e.g., TruthLens, MARE) on the specific axis of stage-gated vs. always-on region features, not only global fusion.

Circularity Check

1 steps flagged

Mild training-data self-reference: Stage-1 annotations are built from the same forensic encoders later injected at examination, undercutting the pure-independence narrative; held-out Acc numbers are not forced by construction.

specific steps
  1. other [Sec. 4.2 Stage 1 / Fig. 3; cf. Sec. 4.1 Eqs. 1–3 stage-gated injection]
    "Multi-signal extraction and scoring: the forensic encoders extract frequency anomaly maps and pixel-level attention maps for each training image; through face parsing masks, pixel-level responses are aggregated into part-level anomaly scores, and the Top-K most suspicious regions are selected as ROIs. ... LLM expert synthesis: the visual descriptions and signal analysis results are jointly fed to an LLM ... into structured five-part annotations."

    The paper’s central independence claim is that part-level forensic evidence is independently sourced and delivered only at examination so planning is unbiased. Yet the supervised trajectories that teach planning and examination are themselves produced by ranking the same spectral/DINOv3 part scores and synthesizing them into the five-part format. The SFT target already encodes which parts the external encoders flag; later stage-gated injection of those same ek vectors can therefore retrieve a memorized signal-to-structure mapping rather than pure mid-chain discovery. This is mild training-time self-reference of the independence narrative, not a by-construction reduction of the held-out Acc numbers.

full rationale

VIGIL is an empirical detection paper, not a closed-form derivation. The load-bearing performance claims (93.1% OmniFake, 92.1% HydraFake) are measured on generators and an external benchmark never used for training or annotation, so they are not predictions forced by fitted inputs. There is no uniqueness theorem, no self-citation chain that forbids alternatives, and no renaming of a known empirical law. The only circular-adjacent element is that Stage-1 SFT targets are synthesized by ranking Top-K ROIs with the same spectral/DINOv3 branches whose embeddings ek are later stage-gated into <part_evidence> (Fig. 3; Eqs. 1–3). That makes the claim of fully independent plan-then-examine reasoning partially self-referential at training time—the model is taught which parts matter using the very signals it later “discovers”—but it does not make the reported accuracy equal to its inputs by definition. The consistency reward’s LLM judge is a training signal only; final metrics remain external Acc. Proportionate score is therefore 2, not 6+.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The paper is an empirical systems contribution; its load-bearing claims rest on a small set of architectural hyper-parameters, standard domain assumptions about facial anatomy and forensic signals, and two invented constructs (the stage-gated injection mechanism and the OmniFake hierarchy). No free parameters are fitted to the final test metrics; the listed scalars are training hyper-parameters chosen once and held fixed.

free parameters (4)
  • reward coefficients λ1, λ2, λ3 = 0.4, 0.4, 0.2
    Set by hand to 0.4 / 0.4 / 0.2 for part-aware, consistency and format rewards; directly control the RL objective that produces the final model.
  • number of anatomical parts K = 8
    Fixed to 8 regions; determines the granularity of planning and evidence aggregation.
  • GRPO group size G and temperature = G=8, temp=1.0
    G=8 rollouts, temperature 1.0; control the variance of the policy gradient used in Stage 3.
  • injection scalars α, γ = zero-init, then learned
    Zero-initialized learnable residual scales for global and part evidence; their learned values are free parameters of the final model.
axioms (4)
  • domain assumption Facial forgery traces concentrate in a fixed set of anatomical regions that can be reliably segmented by an off-the-shelf face parser.
    Invoked in Sec. 4.1 to justify mask-guided pooling and spatial-existence verification; if the parser fails under heavy compression or occlusion the part-level evidence becomes meaningless.
  • domain assumption Frequency-domain high-pass features and DINOv3 pixel features supply forensic signals that are independent of the MLLM’s semantic visual encoder and diagnostic across unseen generators.
    Core premise of the dual-branch forensic encoders (Sec. 4.1); without it the injected evidence is either redundant or non-generalizable.
  • ad hoc to paper An independent LLM judge can recover the correct real/fake label from the model’s evidence text alone, thereby measuring evidence-conclusion coherence.
    Defines the consistency reward R_cons (Eq. 4); the paper treats the judge’s agreement as ground truth for reward shaping.
  • domain assumption Training exclusively on StyleGAN2, SD 1.4 and SimSwap constitutes a realistic minimal-prior forensic scenario from which generalization can be fairly measured.
    Stated in Sec. 3.2 as the design principle of OmniFake; underpins the claim that Level-2–5 results demonstrate true generalization.
invented entities (2)
  • Stage-gated part-level evidence injection no independent evidence
    purpose: Ensures planning is driven only by the model’s own perception while still supplying region-specific forensic signals during examination.
    The gating rule (inject e_k only inside <part_evidence> blocks) is introduced by the paper and has no prior independent experimental validation outside this work.
  • OmniFake 5-Level hierarchical benchmark no independent evidence
    purpose: Provides a progressive generalizability ladder from in-domain to in-the-wild social-media data under a fixed minimal training set.
    Author-constructed dataset; its difficulty ordering and quality-control filters are defined inside the paper and have not been independently audited.

pith-pipeline@v1.1.0-grok45 · 25330 in / 3682 out tokens · 42448 ms · 2026-07-13T20:46:50.248762+00:00 · methodology

0 comments
read the original abstract

Multimodal large language models (MLLMs) offer a promising path toward interpretable deepfake detection by generating textual explanations. However, the reasoning process of current MLLM-based methods combines evidence generation and manipulation localization into a unified step. This combination blurs the boundary between faithful observations and hallucinated explanations, leading to unreliable conclusions. Building on this, we present VIGIL, a part-centric structured forensic framework inspired by expert forensic practice through a plan-then-examine pipeline: the model first plans which facial parts warrant inspection based on global visual cues, then examines each part with independently sourced forensic evidence. A stage-gated injection mechanism delivers part-level forensic evidence only during examination, ensuring that part selection remains driven by the model's own perception rather than biased by external signals. We further propose a progressive three-stage training paradigm whose reinforcement learning stage employs part-aware rewards to enforce anatomical validity and evidence--conclusion coherence. To enable rigorous generalizability evaluation, we construct OmniFake, a hierarchical 5-Level benchmark where the model, trained on only three foundational generators, is progressively tested up to in-the-wild social-media data. Extensive experiments on OmniFake and cross-dataset evaluations demonstrate that VIGIL consistently outperforms both expert detectors and concurrent MLLM-based methods across all generalizability levels.

Figures

Figures reproduced from arXiv: 2603.21526 by Jingjing Chen, Junhao Xu, Xinghan Li.

Figure 1
Figure 1. Figure 1: Overview of the OmniFake dataset. (a) Data collection and quality control pipeline. (b) The hierarchical 5-Level generalization protocol. The model is trained solely on foundational generators and progressively evaluated at each level with increasing difficulty. (c) Performance of existing detectors across levels, showing a widening generalization gap at higher levels. reliability of explanations, subseque… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VIGIL. Left: the part-centric forensic architecture. Specialized forensic en￾coders extract frequency-domain and pixel-level features, which are aggregated into part-level evidence embeddings via face parsing masks. A global evidence summary is injected before reasoning begins; part-level evidence is delivered only during examination through stage-gated injection. Right: the progressive three-s… view at source ↗
Figure 3
Figure 3. Figure 3: Signal-semantic annotation pipeline. Given an input image, forensic encoders extract frequency anomaly maps and pixel-level features, which are aggregated into part-level anomaly scores to identify suspicious regions (Step 1). Multiple off-the-shelf MLLMs independently pro￾duce visual descriptions, with consensus filtering to remove hallucinated observations (Step 2). Finally, an LLM expert synthesizes bot… view at source ↗
Figure 4
Figure 4. Figure 4: Reasoning reversion case. The model initially leans toward “authentic” based on global appearance, but reverses its judgment after examining part-level forensic evidence. The key tran￾sition point is highlighted. 5.4 Further Analyses Reasoning Quality Evaluation. To evaluate reasoning quality, we follow the approaches in [10,58] to conduct two types of assessments: (1) score evaluation using GPT-5.2 and Ge… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detection

    cs.CV 2026-07 conditional novelty 6.0

    Strengthening fine-grained, semantic-anomaly, and pixel-level perception with verifiable rewards, then value-aware on-policy self-distillation, improves generalizable MLLM AI-image detection and adaptation.

  2. SpatialImaginer: Towards Adaptive Visual Imagination for Spatial Reasoning

    cs.CV 2026-04 unverdicted novelty 5.0

    SpatialImaginer integrates visual imagination with textual chain-of-thought to improve spatial reasoning robustness in multimodal large language models.

Reference graph

Works this paper leans on

78 extracted references · 12 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Gpt-image.https://chatgpt.com/images/

  2. [2]

    Midjourney.https://www.midjourney.com/

  3. [3]

    Nanobanana.https://gemini.google/overview/image-generation/

  4. [4]

    Sora2.https://openai.com/index/sora-2/

  5. [5]

    Veo3-1.https://aistudio.google.com/models/veo-3

  6. [6]

    kimi.com/ai-models/kimi-k2-5

    AI, M.: Kimi k2.5: Ai that sees, codes, and works like an expert (2025),https://www. kimi.com/ai-models/kimi-k2-5

    2025

  7. [7]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y ., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

    Brock, A., Donahue, J., Simonyan, K.: Large scale GAN training for high fidelity natural image synthesis. In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net (2019)

    2019

  9. [9]

    arXiv preprint arXiv:2406.09398 (2024)

    Cavia, B., Horwitz, E., Reiss, T., Hoshen, Y .: Real-time deepfake detection in the real-world. arXiv preprint arXiv:2406.09398 (2024)

    Pith/arXiv arXiv 2024

  10. [10]

    In: Forty-first International Conference on Machine Learning (2024)

    Chen, D., Chen, R., Zhang, S., Wang, Y ., Liu, Y ., Zhou, H., Zhang, Q., Wan, Y ., Zhou, P., Sun, L.: Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. In: Forty-first International Conference on Machine Learning (2024)

    2024

  11. [11]

    arXiv preprint arXiv:2504.15271 (2025)

    Chen, G., Li, Z., Wang, S., Jiang, J., Liu, Y ., Lu, L., Huang, D.A., Byeon, W., Le, M., Rin- tamaki, T., et al.: Eagle 2.5: Boosting long-context post-training for frontier vision-language models. arXiv preprint arXiv:2504.15271 (2025)

    arXiv 2025

  12. [12]

    In: Proceedings of the 28th ACM international conference on multimedia

    Chen, R., Chen, X., Ni, B., Ge, Y .: Simswap: An efficient framework for high fidelity face swapping. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2003–2011 (2020)

    2003

  13. [13]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025), https://openreview.net/forum?id=C39ShJwtD5

    Chen, R., Xi, J., Yan, Z., Zhang, K.Y ., Wu, S., Xie, J., Chen, X., Xu, L., Guan, I., Yao, T., Ding, S.: Dual data alignment makes AI-generated image detector easier generalizable. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025), https://openreview.net/forum?id=C39ShJwtD5

    2025

  14. [14]

    arXiv preprint arXiv:2410.06126 (2024)

    Chen, Y ., Yan, Z., Cheng, G., Zhao, K., Lyu, S., Wu, B.: X2-dfd: A framework for explainable and extendable deepfake detection. arXiv preprint arXiv:2410.06126 (2024)

    Pith/arXiv arXiv 2024

  15. [15]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Cheng, S., Lyu, L., Wang, Z., Zhang, X., Sehwag, V .: Co-spy: Combining semantic and pixel features to detect synthetic images by ai. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 13455–13465 (2025)

    2025

  16. [16]

    In: The Thirteenth International Confer- ence on Learning Representations

    Deng, H., Pan, T., Diao, H., Luo, Z., Cui, Y ., Lu, H., Shan, S., Qi, Y ., Wang, X.: Autoregres- sive video generation without vector quantization. In: The Thirteenth International Confer- ence on Learning Representations

  17. [17]

    Advances in neural information processing systems34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  18. [18]

    arXiv preprint arXiv:2006.07397 (2020)

    Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., Ferrer, C.C.: The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397 (2020)

    Pith/arXiv arXiv 2006

  19. [19]

    In: Forty-first international conference on machine learning (2024) 16 X

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y ., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024) 16 X. Li et al

    2024

  20. [20]

    In: Forty-first international conference on machine learning (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y ., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024)

    2024

  21. [21]

    In: International conference on machine learning

    Frank, J., Eisenhofer, T., Schönherr, L., Fischer, A., Kolossa, D., Holz, T.: Leveraging fre- quency analysis for deep fake image recognition. In: International conference on machine learning. pp. 3247–3258. PMLR (2020)

    2020

  22. [22]

    Google: A new era of intelligence with gemini 3 (2025),https://blog.google/ products/gemini/gemini-3

    2025

  23. [23]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Guo, X., Song, X., Zhang, Y ., Liu, X., Liu, X.: Rethinking vision-language model in face forensics: Multi-modal interpretable forged face detector. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 105–116 (2025)

    2025

  24. [24]

    arXiv preprint arXiv:2408.10072 (2024)

    Huang, Z., Xia, B., Lin, Z., Mou, Z., Yang, W., Jia, J.: Ffaa: Multimodal large lan- guage model based explainable open-world face forgery analysis assistant. arXiv preprint arXiv:2408.10072 (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    arXiv preprint arXiv:2505.18660 (2025)

    Huang, Z., Li, T., Li, X., Wen, H., He, Y ., Zhang, J., Fei, H., Yang, X., Huang, X., Peng, B., et al.: So-fake: Benchmarking and explaining social media image forgery detection. arXiv preprint arXiv:2505.18660 (2025)

    arXiv 2025

  26. [26]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Ji, Y ., Hong, Y ., Zhan, J., Chen, H., Lan, J., Zhu, H., Wang, W., Zhang, L., Zhang, J.: Towards explainable fake image detection with multi-modal large language models. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 4398–4407 (2025)

    2025

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jia, S., Lyu, R., Zhao, K., Chen, Y ., Yan, Z., Ju, Y ., Hu, C., Li, X., Wu, B., Lyu, S.: Can chatgpt detect deepfakes? a study of using multimodal large language models for media forensics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4324–4333 (2024)

    2024

  28. [28]

    arXiv preprint arXiv:2601.15624 (2026)

    Jiang, N., Zeng, D., Liu, Y ., Yi, H., Yu, S., Weng, M., Shen, H., Li, Y .: Explainable deepfake detection with rl enhanced self-blended images. arXiv preprint arXiv:2601.15624 (2026)

    arXiv 2026

  29. [29]

    In: European Conference on Com- puter Vision

    Ju, X., Liu, X., Wang, X., Bian, Y ., Shan, Y ., Xu, Q.: Brushnet: A plug-and-play image in- painting model with decomposed dual-branch diffusion. In: European Conference on Com- puter Vision. pp. 150–168. Springer (2024)

    2024

  30. [30]

    In: 2022 IEEE International Conference on Image Processing (ICIP)

    Ju, Y ., Jia, S., Ke, L., Xue, H., Nagano, K., Lyu, S.: Fusing global and local features for gen- eralized ai-synthesized image detection. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3465–3469. IEEE (2022)

    2022

  31. [31]

    In: International Conference on Learning Representations (2018)

    Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. In: International Conference on Learning Representations (2018)

    2018

  32. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversar- ial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401–4410 (2019)

    2019

  33. [33]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110–8119 (2020)

    2020

  34. [34]

    arXiv preprint arXiv:2503.15867 (2025)

    Kundu, R., Jia, S., Mohanty, V ., Balachandran, A., Roy-Chowdhury, A.K.: Truthlens: Visual grounding for universal deepfake reasoning. arXiv preprint arXiv:2503.15867 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Labs, B.F.: Flux.https://github.com/black-forest-labs/flux(2024)

    2024

  36. [36]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, L., Bao, J., Zhang, T., Yang, H., Chen, D., Wen, F., Guo, B.: Face x-ray for more general face forgery detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5001–5010 (2020)

    2020

  37. [37]

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

    Li, O., Cai, J., Hao, Y ., Jiang, X., Hu, Y ., Feng, F.: Improving synthetic image detection towards generalization: An image transformation perspective. In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1. pp. 2405–2414 (2025) VIGIL 17

    2025

  38. [38]

    arXiv preprint arXiv:2503.09314 (2025)

    Li, X., Yu, Y ., Song, X., Shan, H., Chen, J.: Revealing the implicit noise-based imprint of generative models. arXiv preprint arXiv:2503.09314 (2025)

    arXiv 2025

  39. [39]

    arXiv preprint arXiv:2505.19582 (2025)

    Lin, K., Yan, Z., Zhang, K.Y ., Hao, L., Zhou, Y ., Lin, Y ., Li, W., Yao, T., Ding, S., Li, B.: Guard me if you know me: Protecting specific face-identity from deepfakes. arXiv preprint arXiv:2505.19582 (2025)

    arXiv 2025

  40. [40]

    2: Pushing the frontier of open large language models

    Liu, A., Mei, A., Lin, B., Xue, B., Wang, B., Xu, B., Wu, B., Zhang, B., Lin, C., Dong, C., et al.: Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556 (2025)

    Pith/arXiv arXiv 2025

  41. [41]

    In: European Conference on Computer Vision

    Liu, B., Yang, F., Bi, X., Xiao, B., Li, W., Gao, X.: Detecting generated images by real images. In: European Conference on Computer Vision. pp. 95–110. Springer (2022)

    2022

  42. [42]

    Advances in neural information processing systems36, 34892–34916 (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y .J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023)

    2023

  43. [43]

    In: Proceedings of the IEEE international conference on computer vision

    Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730–3738 (2015)

    2015

  44. [44]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ojha, U., Li, Y ., Lee, Y .J.: Towards universal fake image detectors that generalize across generative models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 24480–24489 (2023)

    2023

  45. [45]

    com / index / introducing-gpt-5-2

    OpenAI: Introducing gpt-5.2 (2025),https : / / openai . com / index / introducing-gpt-5-2

    2025

  46. [46]

    arXiv preprint arXiv:2505.02013 (2025)

    Peng, S., Wang, Z., Gao, L., Zhu, X., Zhang, T., Liu, A., Zhang, H., Lei, Z.: Mllm-enhanced face forgery detection: A vision-language fusion solution. arXiv preprint arXiv:2505.02013 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    In: The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rom- bach, R.: SDXL: improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net (2024)

    2024

  48. [48]

    In: European conference on computer vision

    Qian, Y ., Yin, G., Sheng, L., Chen, Z., Shao, J.: Thinking in frequency: Face forgery detection by mining frequency-aware clues. In: European conference on computer vision. pp. 86–103. Springer (2020)

    2020

  49. [49]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image syn- thesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  50. [50]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: Faceforensics++: Learning to detect manipulated facial images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1–11 (2019)

    2019

  51. [51]

    In: ACM SIGGRAPH 2022 conference proceedings

    Sauer, A., Schwarz, K., Geiger, A.: Stylegan-xl: Scaling stylegan to large diverse datasets. In: ACM SIGGRAPH 2022 conference proceedings. pp. 1–10 (2022)

    2022

  52. [52]

    In: Proceedings of the IEEE international conference on computer vision

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam: Vi- sual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618–626 (2017)

    2017

  53. [53]

    arXiv preprint arXiv:2402.03300 (2024)

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y ., Wu, Y ., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    Pith/arXiv arXiv 2024

  54. [54]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shiohara, K., Yamasaki, T.: Detecting deepfakes with self-blended images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18720–18729 (2022)

    2022

  55. [55]

    arXiv preprint arXiv:2508.10104 (2025) 18 X

    Siméoni, O., V o, H.V ., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V ., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025) 18 X. Li et al

    Pith/arXiv arXiv 2025

  56. [56]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Tan, C., Tao, R., Liu, H., Gu, G., Wu, B., Zhao, Y ., Wei, Y .: C2p-clip: Injecting category common prompt in clip to enhance generalization in deepfake detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 7184–7192 (2025)

    2025

  57. [57]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tan, C., Zhao, Y ., Wei, S., Gu, G., Liu, P., Wei, Y .: Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28130–28139 (2024)

    2024

  58. [58]

    In: International Confer- ence on Learning Representations (2026)

    Tan, H., Lan, J., Tan, Z., Liu, A., Song, C., Shi, S., Zhu, H., Wang, W., Wan, J., Lei, Z.: Ver- itas: Generalizable deepfake detection via pattern-aware reasoning. In: International Confer- ence on Learning Representations (2026)

    2026

  59. [59]

    Advances in neural information processing sys- tems37, 84839–84865 (2024)

    Tian, K., Jiang, Y ., Yuan, Z., Peng, B., Wang, L.: Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing sys- tems37, 84839–84865 (2024)

    2024

  60. [60]

    Advances in Neural Information Processing Systems37, 87310–87356 (2024)

    Tong, P., Brown, E., Wu, P., Woo, S., Iyer, A.J.V ., Akula, S.C., Yang, S., Yang, J., Middepogu, M., Wang, Z., et al.: Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems37, 87310–87356 (2024)

    2024

  61. [61]

    Wang, S.Y ., Wang, O., Zhang, R., Owens, A., Efros, A.A.: Cnn-generated images are sur- prisingly easy to spot... for now. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8695–8704 (2020)

    2020

  62. [62]

    In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    Wang, X., Li, Y ., Zhang, H., Shan, Y .: Towards real-world blind face restoration with gen- erative facial prior. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    2021

  63. [63]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(12), 15462–15476 (2023)

    Wang, Z., Zhang, J., Chen, T., Wang, W., Luo, P.: Restoreformer++: Towards real-world blind face restoration from undegraded key-value pairs. IEEE Transactions on Pattern Analysis and Machine Intelligence45(12), 15462–15476 (2023)

    2023

  64. [64]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems

    Wen, S., Ye, J., Feng, P., Kang, H., Wen, Z., Chen, Y ., Wu, J., He, C., Li, W., et al.: Spot the fake: Large multimodal model-based synthetic image detection with artifact explanation. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems

  65. [65]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wu, S., Zhang, W., Xu, L., Jin, S., Wu, Z., Tao, Q., Liu, W., Li, W., Loy, C.C.: Harmonizing visual representations for unified multimodal understanding and generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17739–17750 (2025)

    2025

  66. [66]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Xie, L., Wang, X., Zhang, H., Dong, C., Shan, Y .: Vfhq: A high-quality dataset and bench- mark for video face super-resolution. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 657–666 (2022)

    2022

  67. [67]

    arXiv preprint arXiv:2504.09439 (2025)

    Xu, J., Chen, J., Jiao, Y ., Zhang, J., Tan, Z., Li, H., Jiang, Y .G.: Identity-aware vision- language model for explainable face forgery detection. arXiv preprint arXiv:2504.09439 (2025)

    Pith/arXiv arXiv 2025

  68. [68]

    arXiv preprint arXiv:2601.20433 (2026)

    Xu, W., Lu, W., Luo, X., Zhou, J.: Mare: Multimodal alignment and reinforcement for ex- plainable deepfake detection via vision-language models. arXiv preprint arXiv:2601.20433 (2026)

    arXiv 2026

  69. [69]

    arXiv preprint arXiv:2410.02761 (2024)

    Xu, Z., Zhang, X., Li, R., Tang, Z., Huang, Q., Zhang, J.: Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761 (2024)

    Pith/arXiv arXiv 2024

  70. [70]

    In: The Thirteenth International Conference on Learning Representations

    Yan, S., Li, O., Cai, J., Hao, Y ., Jiang, X., Hu, Y ., Xie, W.: A sanity check for ai-generated image detection. In: The Thirteenth International Conference on Learning Representations

  71. [71]

    Advances in Neural Information Processing Systems37, 29387–29434 (2024)

    Yan, Z., Yao, T., Chen, S., Zhao, Y ., Fu, X., Zhu, J., Luo, D., Wang, C., Ding, S., Wu, Y ., et al.: Df40: Toward next-generation deepfake detection. Advances in Neural Information Processing Systems37, 29387–29434 (2024)

    2024

  72. [72]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yang, Y ., Qian, Z., Zhu, Y ., Russakovsky, O., Wu, Y .: Dˆ 3: scaling up deepfake detection by learning from discrepancy. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 23850–23859 (2025) VIGIL 19

    2025

  73. [73]

    ai/blog/glm-4.6v

    Z.ai: Glm-4.6v: Open source multimodal models with native tool use (2025),https://z. ai/blog/glm-4.6v

    2025

  74. [74]

    In: European conference on computer vision

    Zhang, Y ., Colman, B., Guo, X., Shahriyari, A., Bharaj, G.: Common sense reasoning for deepfake detection. In: European conference on computer vision. pp. 399–415. Springer (2024)

    2024

  75. [75]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zheng, Y ., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y ., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18697–18709 (2022)

    2022

  76. [76]

    Advances in Neural Information Processing Systems35, 30599–30611 (2022)

    Zhou, S., Chan, K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems35, 30599–30611 (2022)

    2022

  77. [77]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Zhou, Z., Luo, Y ., Wu, Y ., Sun, K., Ji, J., Yan, K., Ding, S., Sun, X., Wu, Y ., Ji, R.: Aigi- holmes: Towards explainable and generalizable ai-generated image detection via multimodal large language models. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 18746–18758 (2025)

    2025

  78. [78]

    Advances in Neural Information Processing Systems36(2024)

    Zhu, M., Chen, H., Yan, Q., Huang, X., Lin, G., Li, W., Tu, Z., Hu, H., Hu, J., Wang, Y .: Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems36(2024)

    2024