Face forgery detection methods often perform well on seen manipulations but generalize poorly to unseen ones. Multimodal Large Language Models (MLLMs) can recognize facial parts and assess their visual plausibility. Existing MLLM-based methods typically express these judgments through textual explanations, but seldom treat individual part decisions as explicit learning targets. We present FaceJev, an MLLM-based method for face forgery detection through direct visual decisions. Following the Jev formulation, FaceJev answers eight independent questions about facial parts using a shared image encoding and reads anomaly probabilities from the native answer logits. A face is classified as forged if any part is predicted to be abnormal, directly linking the detection result to the part decisions. We construct PartForensics, a dataset of approximately 160k images with binary anomaly targets for the eight facial parts, to supervise these decisions. We further propose FS-GRPO, which optimizes each part decision through an exact reward comparison with its alternative, using both detection and part rewards. FaceJev outperforms the strongest baseline by 7.25 percentage points on unseen manipulations. Extensive ablation studies validate our design choices and the effectiveness of the main components.
Figures & tables
Figure 1: FaceJev detects forgeries through direct visual decisions. Eight independent questions share an image encoding, with anomaly probabilities read from native answer logits. An image is classified as forged if the maximum part probability reaches 0.5.
Figure 2: Training FaceJev. (a) FS-SFT learns anomaly probabilities from the native answer logits using the binary anomaly targets in PartForensics. (b) FS-GRPO samples binary state vectors and optimizes each part decision through an exact reward comparison with its alternative, using both detection and part rewards. Flipping the mouth preserves the forged image prediction but lowers the part reward, reinforcing the mouth decision, while the incorrect skin anomaly is suppressed. Detection uses only the image and fixed questions.
Figure 3: PartForensics. Top: 159,144 images with binary anomaly targets for eight facial parts, in three splits balanced between real and forged faces. (a) Examples of forged faces from the Seen test set. Boxes and crops are included only to visualize parts with positive targets. (b) The separate Unseen test set covers 17 methods in six categories, and the ring shows the number of methods in each of the categories.
Seen
Unseen
Method
Mean
Swapping
Reenact.
Talking
Synthesis
Editing
Regional
Mean
Avg
No training on this task
Qwen3.5-4B [ 26 ]
51.50
51.34
33.10
49.81
52.84
59.70
49.18
49.33
50.41
Visual Jev [ 41 ]
53.89
52.31
50.02
50.02
56.37
55.51
52.75
52.83
53.36
DeepSeek-V4.1-Flash [ 2 ]
55.65
52.41
52.27
51.38
60.34
57.14
52.88
54.40
55.03
Visual detectors
Table 1: Face forgery detection in BACC (%). Seen Mean weights the nine manipulation methods equally, and Unseen Mean weights the six categories equally. Reenact. denotes face reenactment. Avg is the mean of Seen Mean and Unseen Mean. Bold and underlined values denote the best and second best results, respectively, in each column. FaceJev achieves the highest Mean BACC on both Seen and Unseen.
Configuration
Change
Swapping
Reenact.
Talking
Synthesis
Editing
Regional
Mean
Δ
(a) Learning the decisions
Binary SFT
image label only
73.90
70.05
67.60
83.08
78.52
98.28
78.57
− 2.27
Explanation SFT
label + explanation
75.51
70.87
68.02
82.65
76.20
94.88
78.02
− 2.82
Joint queries
one joint sequence
74.80
69.88
67.73
84.19
73.46
96.24
77.72
− 3.12
No weighting
wi=1
76.60
75.48
76.36
82.05
66.03
94.14
78.44
− 2.40
FS-SFT
independent, weighted
75.78
78.00
76.90
82.14
75.41
96.82
80.84
Table 2: Ablations on Unseen methods in BACC (%). All configurations use Qwen3.5-4B and the same PartForensics training images. Δ is the change of Unseen Mean relative to the last row of each panel, and bold marks the best value for every metric within each panel. FS-SFT achieves the highest Unseen Mean among supervision variants in panel (a), while FS-GRPO leads all six categories in panel (b).
Model / training
Params
Binary
Parts
Δ
Vision encoder with linear heads
ConvNeXt-B [ 18 ]
88M
67.90
64.82
− 3.08
ViT-B/16 [ 4 ]
86M
70.83
67.53
− 3.30
MLLM with direct visual decisions
Qwen3.5-4B, SFT
4B
78.57
80.84
+ 2.27
Qwen3.5-4B, FaceJev
4B
–
84.23
Table 3: Part supervision across backbones. Unseen BACC (%) with image labels (Binary) or eight part targets and the OR readout (Parts). For Qwen3.5-4B, the SFT row reports Binary SFT and FS-SFT, respectively. FaceJev adds FS-GRPO to FS-SFT. Δ is Parts minus Binary, with all differences expressed in percentage points.
Method
Image BACC
Part ACC
Part AUC
Deepfakes
98.09
99.15
99.81
Face2Face
94.83
94.50
97.86
FaceSwap
95.37
96.31
98.91
NeuralTextures
91.20
94.62
96.66
E4S
99.82
77.77
89.73
MAT
99.05
81.83
91.24
Table 4: Image and part decisions of FaceJev on Seen test (%). Part metrics pool all image-part pairs of a method. Mean is the unweighted average across the nine Seen manipulation methods.
Figure 4: Part decisions of FaceJev on eight Seen test images. Labels show predicted anomaly probabilities, and boxes mark the corresponding facial parts in the input face images for illustration.
Method
Output
Throughput
Unseen
Compared methods
FakeVLM [ 38 ]
Score + explanation
2.93
74.09
VERITAS [ 32 ]
Explanation + score
2.39
73.17
SFT ablations
Binary SFT
Binary score
14.60
78.57
Explanation SFT
Score + explanation
1.28
78.02
Table 5: Inference output, throughput, and Unseen BACC (%). We compare FakeVLM and VERITAS with the Binary SFT and Explanation SFT variants in Tab. 2 (a), and the FaceJev method. Throughput (img/s) is measured on one A100 GPU using BF16 and batch size 8. Bold highlights the best result for each metric.
Reliable face forgery detection is critical to the security of online identity verification systems, where missed attacks compromise security and excessive false positives disrupt legitimate users. Specialized forensic detectors achieve strong detection performance but provide limited interpretability, while multimodal large language models (MLLMs) offer strong semantic understanding and interpretable reasoning yet remain substantially weaker for face forgery detection. We argue that a key limitation lies in how visual evidence is acquired: subtle forensic artifacts may be poorly represented at standard resolution, while uniformly processing all cases at higher resolution is computationally inefficient. We therefore introduce ForensicZoom, an industrial-grade MLLM framework for adaptive visual inspection. ForensicZoom first equips a general-purpose MLLM with forensic-aware visual representations and aligns the language model with these features. Its central mechanism, NEED_ZOOM, enables the model to autonomously request magnified views of suspicious regions when the initial evidence is insufficient, turning fixed-pass classification into adaptive multi-round forensic reasoning. The zoom behavior is learned through reward shaping that balances detection accuracy with unnecessary visual inspection, concentrating additional computation on difficult cases. A final attribution optimization stage improves natural-language forensic reports while preserving detection performance. On large-scale industrial identity verification data, ForensicZoom achieves over 97% TPR at 0.1% FPR, substantially outperforming both specialized detectors and existing MLLM-based methods while producing actionable forensic attributions. These results demonstrate that ForensicZoom can provide an effective path toward accurate, interpretable, and scalable MLLM-based face forgery detection.
The rapid increase of deepfakes has raised significant concerns due to their spread on social media. Traditional multi-face forgery detectors crop and verify each face independently, ignoring background context and inter-face relationships, which often yields suboptimal performance. To overcome these limitations, we leverage instruction-based Large Vision-Language Models (LVLMs), which can interpret entire images and follow complex textual instructions. We propose a simple yet effective single-stage multi-face forgery detector, called IMFD (Instruction-based Multi-face Forgery Detector), which is trained end-to-end to jointly localize faces and predict per-face forgery labels. Rather than treating face box prediction only as a joint objective, IMFD explicitly integrates predicted face bounding boxes into the instruction as visual cues that enhance instruction grounding and forgery detection. To support the training and evaluation of IMFD, we convert existing multi-face forgery datasets into an instruction-based format. Experimental results and analyses show that IMFD improves multi-face forgery detection by integrating face bounding boxes into the instruction, and consistently outperforms various state-of-the-art methods.
Dasom Choi, Sangjun Moon, Hyeongchan Im +5
Chungnam National University · Institute of Science Tokyo · Nara Institute of Science and Technology (NAIST)
The rapid evolution of generative paradigms has enabled the creation of highly realistic imagery, which escalates the risks of identity fraud and the dissemination of disinformation. Most existing approaches frame face forgery detection as a fully supervised binary classification problem. Consequently, these models typically exhibit significant performance decay when tasked with detecting forgeries from previously unseen generative paradigms. Furthermore, these methods focus exclusively on either DeepFakes or fully synthesized faces, thereby failing to provide a generalized framework for universal face forgery detection. In this paper, we address this challenge by introducing FADNet (Face Authenticity Detector Net), which reformulates face forgery detection as a one-class classification (OCC) task. By training exclusively on authentic facial data to capture their intrinsic representations, FADNet flags any image whose feature embedding deviates significantly from the learned distribution of real faces as a forgery. In this way, FADNet requires only a single training process without retraining for each specific forgery method. The framework incorporates Evidential Deep Learning (EDL) to quantify predictive uncertainty and utilizes a plug-and-play pseudo-forgery image generator (PFIG) to tighten decision boundaries around authentic data. Extensive experimental evaluations on the DF40 and ASFD benchmarks demonstrate that FADNet achieves superior performance and generalization capabilities. Specifically, FADNet substantially outperforms existing state-of-the-art (SOTA) methods, yielding a remarkable average accuracy of 96.63% and an average precision of 98.83%. Our code is available at https://github.com/hzx111621/Only-Train-Once.
Qingchao Jiang, Zhenxuan Hou, Zhiying Zhu +3
School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China · College of Computer Science and Artificial Intelligence, Fudan University, Shanghai 200433, China · Institute for Infocomm Research (I2R), Agency for Science, Technology and Research (A*STAR), Singapore.