Relational Quality Assessment of AI‐Generated Face‐Swapped Images via Multi‐Image Vision‐Language Reasoning

ABSTRACT

Quality assessment of face‐swapped images is difficult because perceptual degradation is often caused by structured semantic inconsistency rather than by conventional low‐level distortion alone. A swapped face may preserve sharp texture and natural colour statistics, but still be perceived as low quality when identity transfer, facial boundaries, illumination or contextual compatibility is inconsistent across the source, target and swapped images. Most existing no‐reference image quality assessment (NR‐IQA) methods predict quality from an individual image, which makes it difficult for them to capture the relational dependencies among the source, target and swapped images. To address this limitation, we develop a multimodal framework for face‐swapping quality assessment that jointly models semantic consistency and perceptual fidelity. A large vision‐language model is used to analyse the source, target and face‐swapped images jointly, and two complementary prompts are designed to separate degradation attributes from their spatial manifestations. In parallel, a fine‐tuned MANIQA branch extracts low‐level quality‐aware features from the swapped image. The two streams are integrated by a semantic‐perceptual bidirectional fusion (SPBF) module, which performs bidirectional feature‐level interaction between semantic and perceptual representations. Experiments on the MDID benchmark show that the proposed method achieves higher correlation with human opinion scores than representative IQA baselines.