Vontra/detectra-v1

Detectra v3

On-device AI-image detector powering the Detectra Chrome extension: a ViT-S/16 @384 binary classifier (sigmoid(logit) = probability the image is AI-generated), exported to single-file fp16 ONNX for ONNX Runtime Web (WebGPU/WASM).

Fine-tuned from the MIT-licensed Community Forensics ViT-S (Park & Owens, CVPR 2025) on modern generators (DALL·E 3, Midjourney, Flux, SD3.5, Recraft, HiDream), in-the-wild social media (WildRF train), and a replay slice of CommunityForensics-Small — plus, new in v3, non-photographic REAL classes that defeat most detectors: human paintings (WikiArt), hand-drawn anime, meme composites and webpage screenshots. GPT-4o and Ideogram were held out of training entirely.

Preprocessing: shortest edge → 440 (bilinear) → center-crop 384 → [0,1] → ImageNet normalize. Input pixel_values 1×3×384×384 fp32, output logit 1×1.

Held-out results through the extension pipeline @0.65 threshold: WildRF-test 97.0% balanced accuracy (mangled 96.9%), modern-generator eval 99.4% BA (TPR 98.9% incl. never-trained GPT-4o/Ideogram). Per-category TNR on difficult reals: paintings 99.7% (held-out n=1,200), hand-drawn anime 100%, memes 97.6%, webpage screenshots 100%. Full training + evaluation code: https://github.com/ashhart/Detectra

sha256(model.onnx) = 1414b9aafaa01a644ed706224973f09b53a1388282104409862df9893b1b962b

Files

• model.onnx — fp16 ONNX export (deployed by the extension)
• model.safetensors — fp32 fine-tuned checkpoint (for further training/audit)