Multimodal Large Models for Image Tampering Detection and Explanation: From Detection to Reasoning

With the rise of AI-generated content, image tampering has become a growing concern across legal, journalistic, and public safety domains. Existing models often lack interpretability and sensitivity to subtle manipulations. In this work, we propose a lightweight and explainable multimodal framework based on LLaVA-1.5 for image tampering detection and reasoning. By constructing task-specific datasets and reformulating detection as a visual question answering task, our model can identify, describe, and explain manipulations in natural language. We enhance low-level forensic awareness through a dual-branch feature fusion of CLIP semantics and TruFor forensic cues. Experimental results on four benchmarks demonstrate competitive detection accuracy and superior performance in tampering description and rationale explanation. Our method offers a deployable and interpretable solution for real-world content forensics.