Towards Generalizable Deepfake Detection: An Inconsistency-Aware Vision-Language Model Tuning Approach

Deepfakes generated by advanced AI models pose growing challenges to digital media authenticity. Large Vision-Language Models (VLMs) have recently been explored for image forensics due to their multimodal representation ability. However, many existing VLM-based deepfake detection methods keep the visual encoder fixed to preserve pre-trained knowledge, which may limit the model's sensitivity to low-level inconsistency artifacts that are important for deepfake detection. To address this issue, we propose IncoTune, an inconsistency-aware tuning framework that introduces trainable vision-side adaptation into the visual encoder and applies Directional Low-Rank Adaptation (DoRA) to selected linear projection layers in both the visual encoder and the language model. We further report an empirical observation in the ablation study: under the evaluated adapter configuration, replacing LoRA with DoRA in the frozen-vision setting does not improve the average AUC, whereas DoRA provides additional gains when combined with vision-side adaptation. Experimental results on FaceForensics++, DFD, Celeb-DF, DFDC, and DFDCP show that IncoTune improves cross-dataset frame-level detection performance over the frozen-vision baseline and achieves competitive performance among representative frame-level methods, while updating only 27.0M adapter parameters during training. Robustness experiments further evaluate the model behavior under common image degradations.