Loke-60000/slop-detect-vit-s-onnx

Slop Detect ViT-S (int8 ONNX)

Binary AI-generated-image detector powering the Slop Detect Chrome extension. P(ai) = sigmoid(logit); input 3x384x384, CLIP normalization (resize shortest edge to 440, center-crop 384). Confidence threshold 0.65.

Fine-tuned from Community Forensics ViT-S (CVPR 2025, MIT) on ~177k images (CF-Small train subsample across ~2,400 generators plus FLUX / SD3.5 / Midjourney v6 / DALL-E 3 / GPT-4o / Imagen 4 samples; reals from COCO train2017, FFHQ train split, Unsplash, Pexels, Flickr, LAION) with per-sample web degradation (random downscale 25-100%, JPEG q40-98). Calibrated so the balanced-accuracy-optimal operating point sits at the 0.65 displayed confidence; dynamic int8.

Held-out benchmark (21 unseen generators, web degradation): 85.3% balanced accuracy at the 0.65 threshold, measured end-to-end through the extension. Details: RESULTS.md.

sha256(model.onnx) = d431efd677cc124ab68c3f1b20d628b7e9a8362a99803b1878475060b05cff5a