Real-world forgery is inherently heterogeneous: it manifests not only as digital manipulation but also as physical spoofing, and is carried not only by images but also by videos. Despite substantial progress, a unified anti-forgery foundation model covering such a broad threat spectrum remains absent. In this paper, we present Venus-DefakerOne2, which substantially extends the forensic capability boundary along both the forgery-form and media-type dimensions. Trained on tens of millions of open-source and in-house samples, it incorporates two key designs. Tide-Style-Tuning alternates LoRA and full-parameter fine-tuning to balance computational efficiency, generalization, and hard-sample fitting while mitigating catastrophic forgetting. Entropy-based Test Adaptation determines the optimal prediction modality by analyzing prediction entropy. Venus-DefakerOne2 achieves new state-of-the-art performance across a comprehensive suite of benchmarks, including DeepFake faces (96.7%), AI-generated images (93.3%), tampered documents (87.6%), manipulated natural images (89.6%), physical spoofs (93.1%), and AI-generated videos (96.0%). It significantly outperforms both specialized expert models and its predecessor, DefakerOne, establishing itself as a powerful and versatile foundation model for binary-decision forensic tasks.
Figures & tables
Figure 1: DefakerOne2 significantly expands defense coverage along both the forgery-form and media-type dimensions.
Figure 2: DefakerOne2 consistently outperforms its predecessor DeFakerOne and existing state-of-the-art detectors across all image and video forensic domains.
Figure 3: The overall framework of our DefakerOne2.
Figure 4: The proposed Entropy-Based Test Adaptation.
Figure 5: Our data pipeline.
Figure 6: Fake video samples from our Realistic Rumor Video Benchmark (RRVBench), which focuses on real-world rumor scenarios such as military actions, political events, public unrest, health emergencies, etc. They are created by the latest video generators.
Method
DTD
FFDN
Trufor
CDFA
Effort
DRCT
F.MOE
Ours V1
Ours V2
FF-c23
21.2
73.4
53.9
94.2
92.5
66.7
47.7
99.4
99.5
FF-c40
20.8
79.1
49.2
79.7
80.4
59.2
64.9
84.7
88.2
FF-DF
49.8
40.7
62.1
99.1
99.0
78.8
46.0
99.8
99.8
FF-F2F
50.0
43.8
53.8
93.8
92.0
59.3
36.4
99.2
99.3
FF-FS
50.2
42.6
47.0
96.5
97.2
76.5
43.1
99.5
99.6
FF-NT
49.9
49.7
52.8
87.4
81.8
52.5
34.9
98.8
99.1
Table 1: Image-level AUC comparison with forensic vision models on DeepFake detection benchmarks. F.MoE: abbreviation of ForensicMoE, Ours V1: abbreviation of DefakerOne, Ours V2: abbreviation of DefakerOne2
Method
DTD
FFDN
Trufor
CDFA
Effort
DRCT
F.MOE
Ours V1
Ours V2
ForenSynths
48.1
57.3
57.6
49.8
74.8
68.6
71.5
77.2
79.0
D.Forensics
21.8
55.2
51.4
59.7
71.2
72.9
70.9
99.5
99.8
GenImage
43.8
55.1
53.4
59.0
90.5
87.8
94.3
99.7
99.9
Chameleon
36.7
37.9
41.2
55.2
63.1
51.6
58.9
84.7
86.8
Fakeinversion
83.8
38.4
36.7
42.3
84.6
52.4
81.2
91.8
98.0
BFree-Online
36.6
67.8
60.7
38.3
55.9
71.0
32.9
65.7
91.7
Table 3: Image-level accuracy comparison with forensic vision models on AI-generated image detection benchmarks. F.MoE: abbreviation of ForensicMoE, Ours V1: abbreviation of DefakerOne, Ours V2: abbreviation of DefakerOne2
Method
DTD
FFDN
Trufor
CDFA
Effort
DRCT
F.MOE
Ours V1
Ours V2
Doc-FCD
73.2
73.1
68.4
52.3
60.1
77.0
38.9
92.8
99.7
Doc-SCD
69.8
58.7
67.7
39.8
26.7
71.8
24.6
99.8
97.6
Doc-Test
64.3
57.8
60.9
46.9
37.1
66.8
29.8
99.6
98.7
TFR
43.0
51.3
51.8
51.0
50.5
52.2
46.9
99.6
81.8
T-IC13
75.3
61.9
58.8
48.0
19.0
48.9
30.5
54.9
78.5
OSTF
60.3
47.0
59.4
57.7
55.9
57.1
57.9
74.9
91.0
Table 5: Image-level accuracy comparison with forensic vision models on tampered document image detection benchmarks. F.MoE: abbreviation of ForensicMoE, Ours V1: abbreviation of DefakerOne, Ours V2: abbreviation of DefakerOne2
Method
DTD
FFDN
Trufor
CDFA
Effort
DRCT
F.MOE
Ours V1
Ours V2
CASIAv1
46.8
52.5
93.9
63.9
61.9
52.7
7.7
96.3
94.1
COVERAGE
51.5
40.5
72.3
53.6
49.2
52.9
50.6
75.9
80.0
Columbia
62.3
2.20
99.4
88.0
49.2
48.2
81.7
87.7
98.5
NIST16
56.0
38.7
59.4
61.6
73.4
66.7
3.9
85.1
87.3
CocoGlide
51.9
29.6
67.1
67.8
68.2
72.3
19.2
76.0
83.0
Autosplice
42.9
34.3
64.5
77.7
82.0
75.5
37.4
99.9
98.7
Table 7: Image-level AUC comparison with forensic vision models on manipulated natural image detection benchmarks. F.MoE: abbreviation of ForensicMoE, Ours V1: abbreviation of DefakerOne, Ours V2: abbreviation of DefakerOne2
Methods
F. Random
F. Object
F.2-klein
Q.-Edit
S.-Edit
N.B. 2
GPT-I. 2
Avg.
Avg.*
UniFD
50.2
50.8
50.7
50.4
50.7
50.3
50.2
50.5
53.0
LOTA
49.9
49.9
49.9
50.0
49.9
50.0
100.0
57.1
53.2
DRCT
51.0
50.9
50.6
52.2
50.1
57.2
64.0
53.7
60.8
FatFormer
49.9
50.0
50.1
50.1
50.3
50.0
57.0
51.0
-
NPR
45.1
46.0
49.8
55.2
48.8
55.5
80.3
54.4
61.0
SAFE
52.2
51.9
50.0
49.0
48.3
49.5
86.7
55.4
-
Table 9: Average / Balanced accuracy (%) comparison on ManipulationBench. F. Random and F. Object denote FLUX-Fill using random and object masks, respectively. F.2-Klein denotes FLUX.2-Klein. Q.-Edit and S.-Edit are abbreviations for Qwen-Image-Edit and Step1X-Edit-v1p2, respectively. Avg. denotes average performance of official pre-trained models. Avg.* denotes average performance of models officially fine-tuned on ManipulationBench’s training split. Unreported results are denoted by “-”
dataset
FLIP
MVP
LCDF
DTD
FFDN
Trufor
CDFA
Effort
DRCT
F.MoE
Ours V1
Ours V2
CASIA-FASD
98.5
93.8
54.7
25.0
56.2
28.5
45.3
79.0
49.0
25.0
64.3
90.3
DLC2021
55.5
55.5
56.9
44.5
55.3
39.4
34.8
54.7
49.6
44.2
59.6
99.9
GenBuster++
50.5
71.6
49.8
50.0
47.0
51.0
48.8
59.4
49.6
57.8
81.4
88.6
GenBuster-ID
49.1
68.9
51.9
50.0
51.6
50.4
58.0
58.8
71.4
69.7
80.7
99.9
GenBuster-OOD
49.5
77.9
53.3
50.0
44.7
51.7
59.6
54.9
67.7
60.2
83.5
99.6
LOKI
52.5
55.5
55.2
33.6
63.1
35.8
51.1
57.2
63.2
33.6
64.9
92.9
Table 10: Video-level accuracy comparison with forensic vision models on video forgery detection benchmarks. F.MoE: abbreviation of ForensicMoE, Ours V1: abbreviation of DefakerOne, Ours V2: abbreviation of DefakerOne2
Detector
Wan2.2 dyn.
Wan2.2 Light.
LTX
Omni Weav.
Happy Horse
Run way
Kling
SeeDance 2.0
Hai luo
All avg.
Image / Frame-level Forensic Detectors
CNNSpot
35.5
43.8
64.0
45.8
39.2
31.9
45.4
40.4
53.9
44.4
NPR
50.6
55.2
58.6
50.9
54.6
52.3
44.3
48.7
54.6
52.2
UnivFD
32.7
43.0
51.2
39.9
43.0
38.4
39.1
35.0
44.0
40.7
ForgeLens
60.0
69.6
64.1
63.0
64.9
59.5
59.2
48.1
61.7
61.1
Video / Temporal-level Forensic Detectors
Table 14: Video-level AUC comparison of traditional forensic detectors across nine RA-Bench generative subsets
Generator
FLIP
MVP
LCDF
DTD
FFDN
TruFor
CDFA
Effort
DRCT
F.MoE
Ours V1
Ours V2
Happy Horse-1.1
40.0
46.9
55.5
50.0
55.9
49.9
53.6
43.9
71.5
50.2
53.9
96.3
Hunyuan Video-1.5
59.6
44.8
55.1
50.0
49.6
49.6
54.6
48.0
64.1
51.5
50.3
95.9
JoyAI- Echo
39.4
33.8
50.2
50.0
53.1
50.7
52.1
50.2
57.2
51.9
51.9
96.5
KlingV3 Video
51.6
55.7
63.1
50.0
42.6
49.3
60.9
54.3
48.4
51.6
56.2
96.1
KlingV3 Omni
46.9
66.2
63.6
50.0
43.8
49.1
61.7
55.0
47.1
52.1
56.2
96.3
LongCat Video
50.6
58.0
57.1
50.0
57.2
49.4
55.2
49.6
62.2
58.1
58.7
95.6
Table 16: Video-level balanced accuracy comparison of traditional forensic detectors across eleven RRVBench generative subsets
Table 18: Pixel-level forgery localization performance comparison under binary F1 metric. The upper block reports document benchmarks with document-oriented baselines, and the lower block reports nature benchmarks with nature-oriented baselines
Domain
DTD
FFDN
Trufor
CDFA
Effort
DRCT
F.MoE
Ivy-xDetector
Veritas
Ours V1
Ours V2
Image Domain
DeepFake
36.8
57.4
53.4
87.9
88.2
62.4
40.9
74.0
13.6
95.8
96.7
AIGC
44.7
49.5
46.2
49.8
66.1
68.2
57.7
79.9
72.2
87.5
93.3
Document
59.1
56.6
58.5
47.8
41.7
58.8
42.0
40.2
47.7
81.4
87.6
Nature
52.5
34.8
71.5
64.7
63.4
62.5
25.6
51.8
2.0
86.7
89.6
Physical
32.0
60.0
34.2
39.3
67.6
44.6
31.9
35.5
51.2
50.7
87.2
Table 19: Domain-wise summary of full-spectrum forgery detection. ‘Physical’ denotes average accuracy on CASIA-FASD, DLC2021 and SCVD. ‘Digital’ denotes average accuracy on GenBuster-ID, GenBuster-OOD, GenBuster++, LOKI, VideoPhy2.
Settings
All-LoRA
All-Full
Ours (Stage 1)
Ours (Stage 2)
Ours (Stage 3)
Image Average
87.3
87.8
84.6
88.4
90.9
Video Average
93.6
93.0
91.8
94.2
95.7
Table 20: Ablation of the proposed Tide-Style Tuning. Image/video-level detection task
Acc. / Ratio
Frame Acc.
Video Acc.
Frame Ratio
Video Ratio
Ours Acc.
CASIA-FASD
90.0
89.8
61.5%
38.5%
90.3
DLC2021
99.9
99.7
95.4%
4.6%
99.9
GenBuster++
72.2
88.4
23.7%
76.3%
88.6
GenBuster ID
99.9
99.9
92.4%
7.6%
99.9
GenBuster OOD
97.2
98.5
23.5%
76.5%
99.6
RABench
98.2
87.4
44.0%
56.0%
96.3
Table 21: Ablation of the proposed Modality Synergism. Image/video-level detection task
Settings
temperature=1, seed=
seed=42, temperature=
42
1024
4096
0
0.5
1
Image Average
90.9
90.8
91.0
90.9
91.0
90.9
Video Average
95.7
95.7
95.6
95.7
95.5
95.7
Table 22: Ablation study of decoding hyperparameters. Image/video-level detection task
Model
DefakerOne
DefakerOne2
Dataset/ Δ
MFFI
MFFI-Distorted
Δ
MFFI
MFFI-Distorted
Δ
AUC
96.1
76.4
-20.5%
97.3
90.1
-7.4%
Table 23: Robustness evaluation against distortions on MFFI dataset under AUC metric. Image-level detection task
Table 24: Ablation study of the proposed Anomaly Magnifier and MLLM Calibration. Pixel-level forgery localization task under binary F1 metric. The upper block reports document benchmarks and the lower block reports nature benchmarks
Figure 7: t-SNE projection of the last-layer hidden states preceding the LLM classification head ( 2,396 samples, 2,048 dimensions, eight flags). All points are randomly sampled from the test sets, with no filtering by prediction. Marker shape denotes the forensic flag; marker color denotes the ground-truth label and prediction correctness (blue: fake and correct; pink: real and correct; yellow: real but wrong; red: fake but wrong). The five annotated manifolds C0–C4 are quantified in Table 25 .
Cluster
aigc
deepfake
doc
frameaigc
All
C0
15 (4.1%, 86.7%)
–
11 (4.6%, 63.6%)
71 (14.1%, 98.6%)
114 (4.8%, 93.0%)
C1
204 (55.4%, 78.4%)
278 (100.0%, 91.7%)
135 (56.0%, 94.1%)
36 (7.2%, 97.2%)
834 (34.8%, 86.7%)
C2
24 (6.5%, 91.7%)
–
34 (14.1%, 94.1%)
88 (17.5%, 88.6%)
828 (34.6%, 90.0%)
C3
94 (25.5%, 100.0%)
–
42 (17.4%, 92.9%)
277 (55.1%, 98.2%)
492 (20.5%, 95.5%)
C4
31 (8.4%, 77.4%)
–
19 (7.9%, 89.5%)
31 (6.2%, 100.0%)
128 (5.3%, 85.2%)
All
368 (100.0%, 85.1%)
278 (100.0%, 91.7%)
241 (100.0%, 92.1%)
503 (100.0%, 96.6%)
2,396 (100.0%, 89.9%)
Table 25: Per-cluster statistics of the t-SNE projection in Figure 7 , over 2,396 inputs randomly sampled from the eight test sets ( frameaigc / framespoof : frames cut with Ada-FPS from AI-generated videos / from physical presentation-attack videos; videoaigc / videospoof : the same two video families under the native video modality). The two sub-tables split the eight flags into the image-derived ones (a) and the video-derived ones (b); both share the same five clusters. Each cell reports the number of samples of that flag falling into the cluster, its share of all sampled instances of the flag (in %), and the detection accuracy within the cell (in %), i.e. count (share%, acc%) ; the dominant cluster of each flag is bolded, and “–” denotes an empty cell. The bottom row of each sub-table aggregates over all clusters, and the All column reports the cluster size together with the overall cluster accuracy.
Figure 8: Per-type view of the t-SNE projection of Figure 7 : the identical embedding ( 2,396 randomly sampled test inputs, 2,048 dimensions) re-plotted once per forensic flag, with each panel showing only the samples of that flag ( n as titled; marker colors as in Figure 7 ).
Setting
Pre-trained Model
FPR
In-house Test Data1
In-house Test Data2
In-house Test Data3
In-house Test Data4
Avg1
Avg2
(a)
DefakerOne (Ours V1)
0.001
94.11
86.12
81.12
-
87.12
-
(b)
InternVL2-2B
0.001
87.37
73.33
85.94
-
82.21
-
(c)
Qwen3.5-2B
0.001
93.93
87.54
92.77
-
91.41
-
(d)
SwinTrans.
0.001
95.13
81.63
72.59
15.66
83.12
66.25
(e)
DINO-MAC
0.001
93.92
90.47
92.78
79.86
92.39
89.26
(f)
DefakerOne2-0.8B (Ours V2)
0.001
94.92
94.19
95.58
80.61
94.90
91.33
Table 26: Evaluation results on in-house test datasets. Pre-trained models are fine-tuned on the data in-house training data. Compared models include DefakerOne Team (2026) , InternVL2 Chen et al. (2024g) , Qwen3.5-2B Qwen3.5 Team (2026) , SwinTransformer (SwinTrans.) Liu et al. (2021b) , DINO-MAC Qu et al. (2026a) that fine-tuned on the same in-house data.
In recent years, the rapid evolution of generative AI has fundamentally reshaped the paradigm of image forgery, breaking the traditional boundaries between document editing, natural image manipulation, DeepFake generation, and full-image AIGC synthesis. Despite this shift toward unified forgery generation, existing research in Fake Image Detection and Localization (FIDL) remains fragmented. This creates a mismatch between increasingly unified forgery generation mechanisms and the domain-specific detection paradigm. Bridging this mismatch poses two key challenges for FIDL: understanding cross-domain artifacts transfer and interference, and building a high-capacity unified foundation model for joint detection and localization. To address these challenges, we propose DeFakerOne, a data-centric, unified FIDL foundation model integrating InternVL2 and SAM2. DeFakerOne enables simultaneous image-level detection and pixel-level forgery localization across diverse scenarios. Extensive experiments demonstrate that DeFakerOne achieves state-of-the-art performance, outperforming baselines on 39 forgery detection benchmarks and 9 localization benchmarks. Furthermore, the model exhibits superior robustness against real-world perturbations and state-of-the-art generators such as GPT-Image-2. Finally, we provide a systematic analysis of data scaling laws, cross-domain artifacts transfer-interference patterns, the necessity of fine-grained supervision, and the original resolution artifacts preservation, highlighting the design principles for scalable, robust, and unified FIDL.
The rapid advancement of deepfake generation techniques poses significant threats to public safety and causes societal harm through the creation of highly realistic synthetic facial media. While existing detection methods demonstrate limitations in generalizing to emerging forgery patterns, this paper presents Deepfake Forensics Adapter (DFA), a novel dual-stream framework that synergizes vision-language foundation models with targeted forensics analysis. Our approach integrates a pre-trained CLIP model with three core components to achieve specialized deepfake detection by leveraging the powerful general capabilities of CLIP without changing CLIP parameters: 1) A Global Feature Adapter is used to identify global inconsistencies in image content that may indicate forgery, 2) A Local Anomaly Stream enhances the model's ability to perceive local facial forgery cues by explicitly leveraging facial structure priors, and 3) An Interactive Fusion Classifier promotes deep interaction and fusion between global and local features using a transformer encoder. Extensive evaluations of frame-level and video-level benchmarks demonstrate the superior generalization capabilities of DFA, particularly achieving state-of-the-art performance in the challenging DFDC dataset with frame-level AUC/EER of 0.816/0.256 and video-level AUC/EER of 0.836/0.251, representing a 4.8% video AUC improvement over previous methods. Our framework not only demonstrates state-of-the-art performance, but also points out a feasible and effective direction for developing a robust deepfake detection system with enhanced generalization capabilities against the evolving deepfake threats. Our code is available at https://github.com/Liao330/DFA.git
Jianfeng Liao, Yichen Wei, Raymond Chan Ching Bon +3
Shenzhen Technology University, Guangdong, China · Singapore Institute of Technology, Singapore, Singapore · The University of Hong Kong, Hong Kong, China +1
Rapid advancements in video diffusion models and temporal editing tools have enabled the generation of highly realistic human-centric videos, posing unprecedented challenges to digital content forensics. Existing benchmarks primarily focus on either face-swapping or global text-to-video synthesis, overlooking the crucial dimensions of human-object or human-human interactions and multi-modal alignment. To address these limitations, we introduce HumanForge, a unified, large-scale, and multi-paradigm human-centric video forgery dataset. To construct and annotate this dataset without labor-intensive manual labeling or hallucinated monolithic prompts, we propose Gen2Anno, a modular active multi-agent pipeline built on LangGraph. Gen2Anno coordinates six specialized agents-ranging from source profiling to MoE-based reference analysis and closed-loop forensic verification-to generate over 18K high-fidelity video segments and produce structured, contrastive omni-annotations containing binary decisions, fine-grained artifact categories, and spatio-temporal localization. Extensive benchmarks using state-of-the-art traditional detectors and Large Multimodal Models (LMMs) demonstrate the significant challenges of zero-shot generalization and fine-grained reasoning on HumanForge. Code and dataset will be publicly released.
Wenbo Xu, Zhimin Chen, Xiaojie Liang +2
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou 510006, China