IC-480Pre-trained M-LLMs (GPT-4o, LLaVA-v1.6-34B, InternVL2-26B, Qwen2-VL-7B) produce imprecise tampering explanations when artifacts require fine-grained pixel-level analysis such as lighting or perspective inconsistencies
The paper evaluates four released M-LLMs on their ability to generate textual explanations of image tampering, measured by cosine semantic similarity (CSS) against GPT-4o-generated ground-truth descriptions across nine datasets spanning photoshop, deepfake, and AIGC-editing. All four models achieve CSS scores between 0.4887 and 0.6760, substantially below the fine-tuned FakeShield (0.7537–0.8873). The authors note that these M-LLMs can leverage pre-training knowledge to make reasonable judgments when tampering causes clear physical-law violations, but they struggle with more precise analyses like detecting lighting or perspective inconsistencies, which reduces overall explanation accuracy.
Ground-truth descriptions were generated by GPT-4o itself, so GPT-4o's CSS is a self-consistency measure rather than an independent reference; the paper does not report per-artifact-type breakdowns of where each model fails.