基于 CLIP 的合成图像检测：理解与评估预测线索

Recent advancements in generative models have led to the production of near-photorealistic synthetic images, posing challenges to the trustworthiness of photographs. As a result, synthetic image detection (SID) has emerged as a crucial area of research. Previous studies have highlighted the differences between synthetic images and real photographs; however, existing SID methods often struggle to generalize to novel generative models and perform poorly in practical applications. CLIP, a foundational vision-language model that generates semantically rich image-text embeddings, demonstrates strong accuracy and generalization capabilities for SID tasks. Nevertheless, the underlying relevant cues embedded in CLIP features remain unclear. It is uncertain whether CLIP-based detectors are merely identifying strong visual artifacts or leveraging subtle semantic biases.