Face-Swap & Deepfake Check
Free · No sign-up · Different privacy model — read before use: this one analysis is designed to run server-side. Nothing is sent anywhere until you press Analyze, and we store nothing.
When the engine is live, pressing Analyze will send ONLY the shown frame to our analysis provider (Sightengine) over an encrypted connection, for this one check, and we will not store the frame or the result. Until then, the button uploads nothing.
What this tool cannot tell you
- It cannot prove a face is real or swapped. Detector scores are model probabilities, and every public detector has meaningful error rates in both directions.
- A low score on a compressed, filtered or tiny face means "not detected in this frame" — nothing more.
- It checks one frame. A video can mix real and swapped segments.
- Detection accuracy collapses after recompression. In one published benchmark, a detector's accuracy fell from 95.3% to 0.2% after the media was recompressed at JPEG quality 50 — and almost everything shared online has been recompressed. Scores on re-uploaded, forwarded or screenshotted media are weak evidence, however confident the number looks.
- Each check costs multiple operations on the provider side. When this checker activates, free checks will be capped per visitor per day so it stays sustainable — no watermarks and no sign-up, just a daily cap.
How face-swap detection works
A face-swap detector is a classifier: a model trained on large sets of real faces and known face-swap outputs, which learns the statistical traces synthesis leaves behind. The traces it weighs include blending boundaries where a generated face was composited onto a scene, skin texture and lighting that don't match between the face and its surroundings, warping artefacts around the jaw and hairline, and frequency-domain patterns that generators tend to produce. Given a frame, the model returns a probability that the face carries those traces — a score about pixels, in one frame, from one model family. It is not a judgement about the person in the image, and it cannot be one: the same traces appear in honest footage that has been filtered, beautified, compressed or shot in poor light.
That fragility is exactly why this checker is not live yet. The analysis cannot run in a browser tab — it needs a dedicated server-side engine, and every check costs real operations on that provider's pricing. Connecting it means choosing the engine carefully and paying for it sustainably, because a cheap or careless detector would produce confident-looking scores that mean very little. Until that is in place, this page explains the method and the limits — and the fully-localEXIF & Provenance Viewer already checks what a file declares about itself, without anything leaving your device.
Frequently asked questions
What does a face-swap check actually look for?
Statistical traces of face synthesis and replacement: blending boundaries, inconsistent skin texture and lighting between the face and the scene, warping artifacts around the jaw and hairline, and generator fingerprints. A dedicated engine (this page is built around Sightengine's deepfake models, not yet connected) scores a face image against patterns learned from known face-swap systems.
Can it prove a video is a deepfake?
No detector can prove that. Scores are probabilities from one model family; heavy compression, beauty filters, low light and aggressive editing all move them, and the best modern swaps defeat public detectors regularly. Treat any result — from any tool, including ours — as one piece of evidence, never a verdict about a person.
Why does the check need my image to leave my device?
Unlike our other tools, this analysis can't run in a browser tab — the models are server-side. So this page is designed to work differently and says so up front: nothing is sent until you press the analyze button, the frame then goes to the analysis provider (Sightengine) over an encrypted connection, and we store nothing. While the engine is not yet connected, nothing leaves your device at all. If you'd rather nothing ever does, the EXIF & Provenance Viewer checks what a file declares about itself, locally.
Which frame should I check?
A clear, front-facing, well-lit frame where the face is large. Profile angles, tiny faces and motion blur produce unreliable scores on every detector. For video, this page grabs a frame at the point you choose.
What does it cost?
When it activates, checks will be free for creators here. Behind the scenes each analysis costs real operations on the provider's pricing, which is why this checker will have fair-use limits instead of watermarks or sign-ups.
Can I check someone else's video?
You can check a frame you lawfully have. Remember the result is a probability about pixels — publishing 'this person is fake' claims from a detector score can be defamation when the score is wrong. Use it to decide what deserves a closer look, not to accuse.
Related tools
- EXIF & Provenance Viewer — the fully-local evidence check (nothing ever leaves your device).
- AI Label Checker — will platforms label this content?
CreatorChecks is not affiliated with YouTube, Google, TikTok, Meta or Sightengine.
Was this result accurate?
Thanks — feedback like this helps us catch errors.
Could not send — please try again.