WH-2 detects synthetic images, videos, and identity assets before they turn into approved accounts, paid claims, fake listings, or reimbursed fraud. Built for teams that need to know whether visual evidence is real — fast, clearly, and at scale.
Synthetic media doesn't replace your KYC, claims, or trust & safety stack — it shows up inside it. WH-2 sits at the asset-check step.
One REST endpoint and a webhook. Drop WH-2 into the same pipeline you already use for identity, fraud, and content checks — no rebuild.
Synthetic media also drives review queues, escalations, chargebacks, and friction for the customers you actually want.
Send the asset, get a verdict, act on it. WH-2 fits next to your existing KYC, claims, or trust & safety stack — no rewrite required.
POST an image or video to /v1/detect. Same shape for every modality.
Pixel + frequency + generator-fingerprint passes run in parallel. Sub-5 ms verdict.
Verdict, confidence, suspect regions, and a short explanation — ready to route.
Designed to generalize to new generators, not memorize the old ones.
The point isn't the score on tests we've seen — it's that WH-2 is built to generalize to generators it hasn't.
Send up to 200 verification samples and we'll come back with a report in 24 hours — no commitment, no signup. Or try it yourself on your own image.
Fully AI-generated images and video, frame-level manipulations, face swaps, and synthetic identity assets — across the major open and commercial generators in circulation.
WH-2 is trained for generalization, not memorization. On benchmarks where we deliberately withhold a generator family during training (ProGAN, for example), it still scores 99%+. New models that show up in the wild are added to our internal evaluation set; you don't have to wait for a release.
Under 5 ms inference per image on our endpoint. End-to-end round-trip from the closest region is typically 25–60 ms.
A verdict (real / fake / tampered), a calibrated confidence score, the suspect regions highlighted, and a short text explanation of what triggered the flag — so a human reviewer can act on it.
