Image Deepfake Detection (IDD) separates manipulated images from authentic ones by spotting artifacts of synthesis or tampering. Although large vision-language models (LVLMs) offer strong image understanding, adapting them to IDD often demands costly fine-tuning and generalizes poorly to diverse, evolving manipulations. We propose the Semantic Consistent Evidence Pack (SCEP), a training-free LVLM framework that replaces whole-image inference with evidence-driven reasoning. SCEP mines a compact set of suspicious patch tokens that best reveal manipulation cues. It uses the vision encoder's CLS token as a global reference, clusters patch features into coherent groups, and scores patches with a fused metric combining CLS-guided semantic mismatch with frequency-and noise-based anomalies. To cover dispersed traces and avoid redundancy, SCEP samples a few high-confidence patches per cluster and applies grid-based NMS, producing an evidence pack that conditions a frozen LVLM for prediction. Experiments on diverse benchmarks show SCEP outperforms strong baselines without LVLM fine-tuning.
Figures & tables
Fig. 1: Comparison of LVLM inference paradigms for IDD. (A) Direct inference uses the full image. (B) VQA-style inference formulates IDD. (C) Logits-based inference uses token probabilities. (D) A training-free, evidence-driven framework that selects compact evidence tokens with frequency/noise cues and grid-based NMS for LVLM prediction.
Fig. 2: Overview of SCEP for training-free IDD. (1) Dual-cue patch anomalies are scored in semantic space: frequency is measured by JSD between each patch’s DCT spectrum, while noise is quantified from residuals via the median and MAD. (2) Patch embeddings are clustered around the CLS anchor to select top-k evidence tokens, and grid-based NMS removes redundant ones. The resulting evidence pack conditions a frozen LVLM to predict Real or Fake.
TABLE I: Performance on DFBench across AI-generated subsets. The table compares multiple frozen LVLM backbones and token-pruning baselines, A-ViT and DynamicViT. Rows shaded in lavender indicate token-pruning baselines, while the row shaded in light gray highlights SCEP.
Multimodal large language models (MLLMs) can explain deepfake verdicts in natural language, but such explanations are not necessarily visually grounded in the visual evidence underlying the prediction. A model may describe plausible artifacts inferred from language priors rather than from image evidence. Existing grounding methods improve visual reliance through decoding or attention interventions, but they generally strengthen grounding over the entire image, making them ill-suited for forensic artifacts that are subtle, spatially localized, and image-dependent. We propose Look Before You Judge, a training-free framework that formulates explainable deepfake detection as a sequential evidence acquisition process. Instead of directly predicting image authenticity from holistic visual reasoning, our framework first identifies image-specific candidate evidence regions by contrasting the MLLM's decoder-to-visual attention between an original image and its Gaussian-blurred counterpart. The identified regions are then inspected individually, and the resulting local evidence is integrated with the global image context before reaching a final verdict. The framework operates without manipulation masks, external forensic models, or parameter updates, making it directly applicable to off-the-shelf MLLMs. Across five open-source MLLMs on TriDF and MMTD-Set, our framework improves detection accuracy by up to 12.8%, reduces CHAIR by up to 33.4% and hallucination rate by up to 21.3%, and outperforms representative training-free decoding and attention methods.
Chia-Ling Chen, Yu-Ting Ta, Jian-Yu Jiang-Lin +8
National Taiwan University · National Tsing Hua University · National Yang Ming Chiao Tung University
Recent Deepfake Video Detection (DFD) studies have demonstrated that pre-trained Vision-Language Models (VLMs) such as CLIP exhibit strong generalization capabilities in detecting artifacts across different identities. However, existing approaches focus on leveraging visual features only, overlooking their most distinctive strength -- the rich vision-language semantics embedded in the latent space. We propose VLAForge, a novel DFD framework that unleashes the potential of such cross-modal semantics to enhance model's discriminability in deepfake detection. This work i) enhances the visual perception of VLM through a ForgePerceiver, which acts as an independent learner to capture diverse, subtle forgery cues both granularly and holistically, while preserving the pretrained Vision-Language Alignment (VLA) knowledge, and ii) provides a complementary discriminative cue -- Identity-Aware VLA score, derived by coupling cross-modal semantics with the forgery cues learned by ForgePerceiver. Notably, the VLA score is augmented by an identity prior-informed text prompting to capture authenticity cues tailored to each identity, thereby enabling more discriminative cross-modal semantics. Comprehensive experiments on video DFD benchmarks, including classical face-swapping forgeries and recent full-face generation forgeries, demonstrate that our VLAForge substantially outperforms state-of-the-art methods at both frame and video levels. Code is available at https://github.com/mala-lab/VLAForge.
Jiawen Zhu, Yunqi Miao, Xueyi Zhang +2
Singapore Management University, Singapore · The University of Warwick, UK · Nanyang Technological University, Singapore +1
Forensic deepfake analysis demands more than binary classification: investigators need region-grounded natural language explanations they can verify against the image. Multimodal large language models (MLLMs) are a natural fit, but pretrained MLLMs fail systematically, producing globally coherent text that misses the small localized cues defining manipulations. We argue this is an inductive bias problem rather than a capacity issue: the image-text contrastive objective training MLLM visual encoders optimizes for whole-image semantic summaries, not patch-level forensic detail. The same mismatch explains why prior deepfake reasoning methods target either face manipulation or fully AI-generated content, never both. We propose FORGE, which addresses the mismatch by routing a second visual stream into the language model from a Vision-Only Model (VOM) trained on dense patch prediction rather than image-text alignment. The MLLM's native encoder and the VOM operate on a shared patch grid, which lets us interleave their tokens with preserved spatial correspondence; we show this beats naive concatenation. A two-stage adapter training protocol (generic image-caption alignment, then joint task-specific optimization) prevents the localized stream from overfitting to training-domain manipulations. Across face-manipulated and fully synthetic content, FORGE produces region-referential explanations answering fine-grained attribute queries ("Does the eyes/nose/mouth look real or fake?") and substantially outperforms in-domain baselines on cross-domain evaluations; region-specific evaluation and human studies confirm explanation faithfulness.