ResNet-18 (fine-tuned by me). Started from ImageNet weights and fine-tuned on a public real-vs-AI image set, with JPEG compression added during training. It calls an image AI at 0.61 or higher, a cut-off chosen on my validation images.
SwinV2 (frozen baseline). A pretrained detector (haywoodsloan/ai-image-detector-deploy), used as-is. It says "likely AI" only above 0.99 and "no strong AI signal" below 0.10.
Both models run on a cloud GPU. Images are analyzed in memory and never stored.