Synthetic image detection with CLIP. Understanding and assessing predictive cues

Abstract

Recent generative models produce near-photorealistic images, challenging the trustworthiness of photographs. Synthetic image detection (SID) methods, however, often struggle to generalize across datasets and generative models. CLIP, which embeds images and text in a shared seamantic space, performs well at SID, but the cues underlying its decisions remain poorly understood. We therefore study CLIP-based SID as an empirical interpretability problem rather than proposing a new detector. We introduce SynthCLIC , which pairs real photographs with caption-matched, high-quality diffusion-generated counterparts. We evaluate CLIP-based detectors on SynthCLIC , a GAN-heavy benchmark, and a broad external benchmark, and compare them with a low-level forensic CNN, a broad-generator detector, and a text-grounded concept model. CLIP-based linear detectors reach 0.96 mAP on the GAN-heavy benchmark but 0.92 on SynthCLIC , while cross-family transfer to CNNSpot falls to 0.42 mAP. Within-class associations between detector scores and text-derived cue scores show that higher synthetic scores correspond to cleaner, more compositionally controlled, and technically polished images, whereas lower scores correspond to messier capture conditions and provenance cues characteristic of real photographs. These associations are distributed across many overlapping cues, and their profiles differ strongly across training datasets. CLIP-based and forensic detectors therefore fail in different ways and provide complementary evidence, while broad generator coverage appears important for robust SID.