AI Image Detector — Check if a Photo is AI-Generated, Free

Free online AI image detector: check if a photo is AI-generated with an on-device neural detector + C2PA/metadata forensics. No upload — the image never leaves your device.

🔒 Runs entirely in your browser — your files never leave your device.

About AI Image Detector

Is that image real or AI-generated? This tool combines a neural detector (a Vision Transformer trained to spot generative artifacts) with metadata forensics (C2PA manifests, Stable Diffusion/ComfyUI parameters, generator EXIF/XMP tags) and gives you a calibrated probability — entirely on your device. The image is never uploaded, which matters: detection is exactly the kind of task you don't want to send to a third party.

How to use AI Image Detector

  1. Drop an image into the upload area (JPG, PNG, WebP…; the first visit downloads the ~44 MB detector model once, then it is cached).
  2. Wait a second or two while the detector analyzes the image locally.
  3. Read the verdict: an AI-probability meter, plus any hard evidence found in the file's metadata.

How AI image detection works here

Two layers of evidence

  • Neural analysis: a Vision Transformer (ViT-S/16) trained on millions of real and AI-generated images looks for the subtle artifacts that generators leave behind — noise patterns, frequency signatures, upscaling traces. It outputs a logit, which is Platt-calibrated into an honest probability.
  • Metadata forensics: the file itself is scanned for proof of provenance — C2PA manifests (the emerging signed-provenance standard), Stable Diffusion 'parameters' chunks, ComfyUI workflows, and EXIF/XMP tags naming generators like Midjourney, DALL·E, or Firefly.

The two layers combine asymmetrically by design: hard metadata evidence can only RAISE the score (an embedded SD prompt proves AI generation), while its absence proves nothing — most AI images carry no metadata at all. Camera EXIF is shown as context only, because EXIF is trivially spoofable.

Reading the score

The verdict uses a 65% threshold: above it, the image is 'likely AI-generated'; below, 'likely not'. Treat the score as evidence, not certainty — no detector is perfect, especially as generators improve. Very high scores backed by metadata evidence are the most trustworthy; borderline scores (40-70%) are genuinely uncertain.

The tool shows the calibrated neural score and inference time at the bottom of the evidence panel, so you can see exactly what drove the verdict.

Why run detection locally?

Detection tools that ask you to upload the image create a paradox: you're trying to verify authenticity, and you hand the image to an unknown server. Here the model runs on your device via WebAssembly — the same engine family as our other local-AI tools — so the image under scrutiny never leaves your computer. That also makes it safe to check sensitive or private images.

Credits

Detection model: Detectra v1 by Vontra (MIT), inference pipeline from the Detectra project by ashhart (MIT). The model is downloaded once from Hugging Face and cached in your browser.

Frequently asked questions

Is this AI image detector really free?

Yes — free with no account, no watermark, and no limit on the number of images.

Is my image uploaded to a server?

No. The detector model runs locally on your device via WebAssembly. Only the model itself is downloaded (once, ~44 MB) from Hugging Face; your image never leaves your computer.

How accurate is it?

The detector is a ViT-S/16 model trained on real vs. AI-generated images, with Platt calibration for honest probabilities. It works well on common generators (Midjourney, DALL·E, Stable Diffusion, Flux…) but no detector is perfect — heavily edited images, very old generators, or screenshots can fool it. Metadata evidence (when found) is the strongest signal of all.

Which image formats work?

Anything your browser can decode: JPG, PNG, WebP, AVIF, GIF (first frame), BMP. Metadata forensics is deepest for PNG and JPEG (where SD parameters, C2PA, and EXIF live).

Can AI images pass as real?

Sometimes. Stripping metadata, re-encoding, or screenshots remove most forensic traces, leaving only the neural score — which is strong but not infallible. A 'likely not AI' verdict is evidence, not proof.

Which browsers are supported?

Any modern browser on desktop or mobile: Chrome, Firefox, Safari, and Edge.

Related tools