Deepfake Detector

Check a picture for a swapped or generated face. A model trained on 4,803 generators reads it on your own device.

Checking whether the detector is available…
Three steps

How a deepfake check actually runs

There is no upload and no queue. Nothing downloads because you visited. The model is fetched into your tab on the first check and cached after that, and the picture is decoded by your own browser.

01Hand it a picture

Drag a file in or click to browse. One picture at a time here, because this model is heavier than the one on the home page and reads a single frame closely.

02The frame is read as a whole

The picture is centre-cropped and normalised the way the model was trained, then scored once. What comes back is a likelihood that generation touched this frame.

03You get a number and the caveats

A 0 to 100 figure on the same five bands the picture detector uses, plus what was measured about the file that should temper your reading of it.

Try it on a picture
Three problems

The three things people mean by the word

“Deepfake” covers three things that are made differently and have to be found differently. A tool that catches one is not thereby a tool that catches the others, so here is which is which, and which of them this detector reads.

Whole frame
94 typical score

Generated outright

No camera involved. A model produced the entire picture from a prompt, including a person who does not exist. This is the case the detector reads most confidently.

One region
71 typical score

A face swapped in

The room, the clothes and the light are genuine; someone else's face has been fitted onto the body. A small patch of generated pixels inside a real photograph.

One region, moving
66 typical score

A face altered

The same person, changed: expression edited, mouth reshaped to fit words they never said. In a still this carries the same local signal a swap does; across a clip it is harder.

In the wild

What people bring to a deepfake check

Almost every picture that reaches this page is of a real person. That is the reason the file never leaves your device, and the reason the limits are stated rather than buried.

  • Social profiles

    A generated face on a borrowed body, used to open accounts and start conversations.

  • Non-consensual imagery

    Face swaps and nudification used against someone who never agreed to any of it.

  • Verification selfies

    A synthetic selfie presented to an automated identity check.

  • Executive impersonation

    A photograph of a director attached to an instruction they never gave.

  • Dating and romance

    A face that passes every video call because it was never a face.

  • Political imagery

    A candidate placed somewhere they never stood, timed before a vote.

  • Press photographs

    A picture offered to a newsroom on a deadline too short to check it.

  • Staff directories

    A headshot for an employee who exists only on the company website.

  • Extortion threads

    An image built from public photographs, then used as a threat.

  • Dispute evidence

    A photograph submitted to a tribunal to show something that never happened.

  • Missing person appeals

    An appeal that spreads faster than anyone can check whether the person exists.

  • Celebrity endorsements

    A well-known face endorsing an investment they have never heard of.

Reading the number

What the score says, and how firmly

What comes back is a likelihood from 0 to 100 on the same five bands the picture detector uses, with the decision line at 65. Sharing the scale is deliberate: somebody who has learned to read a 61 on the home page should not have to learn a second one here.

Read the middle band as the middle band. A swapped face is a small patch of generated pixels inside a genuine photograph, and a score that lands at 60 on such a picture is not the detector failing — it is an honest summary of an image that is mostly real.

The five bands, what each one means for a picture and what to do about it
Score Band What it means What to do next
0 – 20 No manipulation signal Nothing in the pixels points towards generation or a swap. Nothing in the pixels points to a generated or swapped face. Check where it came from anyway.
20 – 45 Probably authentic A faint reading, of a strength ordinary editing produces. A faint reading. Ordinary retouching and compression sit here too.
45 – 65 Inconclusive The picture sits between the two populations the model separates. The honest middle. Screenshots and heavy compression collect here — find a better copy.
65 – 90 Likely manipulated Over the decision line, in the region where edited and generated frames land. Over the line, but near enough that resizing could have lifted it. Corroborate before acting.
90 – 100 Manipulated As firm as this analysis gets on a single frame. A strong reading. Work on the basis that the picture was generated or altered.

The bands are fixed and published, so a 72 today means what a 72 meant last month.

A score is a likelihood, not proof, and a picture of a real person is exactly the case where that distinction matters most.

Under the hood

What the model is actually looking at

It reads how the pixels are put together, not who is in the picture. Generation and face swapping both reconstruct a region of the image, and reconstruction leaves statistical residue: noise that behaves too evenly across a surface, edges that resolve differently from optical ones, texture that repeats at scales a lens and a sensor do not produce together.

It was trained on 2.7 million pictures drawn from 4,803 distinct generators. That breadth is the point. A detector trained on three or four popular models learns those models; one trained across thousands learns something closer to the general shape of reconstruction, which is what gives it any chance at all against a generator released after it was built.

For a swapped face, the readout deliberately keeps the middle of the scale available. Scored as one image, a swap is outnumbered by the genuine room, clothing and light around it, and a tool that pushed such a picture to 95 would be lying about how much of the frame it actually found.

None of this reads minds or reconstructs history. It measures how closely a frame resembles the output of a generative process, and a picture that has been screenshotted, resized and reposted has had much of that evidence removed before it reached you.

What is measured, and what is not

  • Measured: pixel statistics across the whole frame, at the model's own crop
  • Measured: file size, dimensions and format, reported as caveats beside the score
  • Not measured: who is in the picture, or whether a face matches a named person
  • Not measured: which generator or face-swap tool was used
  • Not measured: video, motion or lip-sync timing — the video detector samples frames instead
Accuracy

Where it holds up, and where it slips

The model's published evaluation covers 4,803 generators and 2.7 million images, which is the widest training coverage of any detector deployed on this site.

Wide coverage is not the same as an independent out-of-domain evaluation, and none exists for this model. It is deployed at full precision rather than quantised, because quantising it moved individual scores by as much as 97 points in testing, which would have made the published bands meaningless.

What pushes a result towards the middle

  • Screenshots, which resample the whole frame and discard the fine texture
  • Repeated reposting, where each platform re-encodes what the last one produced
  • A small swapped region inside a large genuine photograph
  • Heavy retouching, beauty filters and portrait modes on a real photograph
  • Low resolution, or an image upscaled after the fact
  • Print scans and photographs of screens, which add a second capture

A false positive is most often a heavily retouched photograph of a real person. Portrait modes, skin smoothing and aggressive sharpening all rearrange texture in the direction the model reads as reconstruction, which is why the caveats beside the score name what was measured about the file.

A false negative is most often a swap that survived compression better than the evidence did. If a picture matters and the copy you have is a screenshot, finding the original file is worth more than running the check again.

Read the result as evidence, never as a verdict. A picture of a real person can end a career, and no detector output should be the only thing standing behind that.

What it will not do

  • Tell you who is in the picture, or match a face to a person
  • Name the tool that made the swap
  • Prove a picture is authentic — a low score is the absence of a signal, not evidence of a camera
  • Track a face through video frames or measure lip-sync timing; the video detector samples visual frames instead
  • Distinguish a swap from a wholly generated frame with certainty on a single number
  • Stand on its own in a legal, employment or disciplinary decision

What still beats a detector

Provenance outranks any score. Find the earliest posting of the picture, ask for the file the camera produced rather than a forward, run a reverse image search on the face, and check whether the person shown has said anything about it. For a swap, the original photograph the body came from often still exists somewhere.

What you get

Everything here, at no cost, with nothing held back

There is no paid tier of this deepfake detector holding back the part you need. The score, the band, the measured caveats and the model's coverage are all in the free version, because there is no other version.

A model built for faces

Trained across 4,803 generators rather than the handful a demo detector recognises, which is what a new generator tests.

The same scale as the rest

Five published bands with the decision line at 65, shared with the picture and voice detectors so one reading skill covers all three.

Full precision, deliberately

Nothing is compressed to save bandwidth. A smaller build was tested and moved some scores by nearly the whole scale.

Caveats from the file itself

Screenshots, upscaling and hard compression are detected and named beside the score rather than silently ignored.

Nothing uploaded

A picture of somebody's face is the last thing that should be sent to a server. It is decoded and scored in your own browser.

Saved on your device

Results are kept in your own browser storage so you can come back to them, and clearing your site data removes them.

Formats, limits and requirements

Formats read
JPG, PNG, WebP, AVIF, GIF, BMP
Per check
One picture at a time
Per check
One picture, read as a whole frame
Training coverage
2.7 million images from 4,803 generators
Coverage
Generated frames, swapped faces and altered faces in stills
Scale
0–100 over five published bands, decision line at 65
Runs on
Your device, via WebAssembly. Nothing is transmitted
Privacy

Where the picture goes: nowhere

This is the part where most deepfake detection sites ask you to trust a policy. There is no policy to trust here, because there is no transfer. It matters more on this page than on any other: the pictures people bring to a deepfake check are usually of a real person, often one who has not consented to anything, and sometimes one who is already the victim of the image.

Chosen The page reads the file straight off your device.
Scored The picture is decoded and the model runs in this tab.
Dropped Close the tab and the picture and the decoded pixels are gone.

Read the privacy policy

By eye

What to look at before you even run a check

A detector is the reliable path, but a swap has to survive a boundary between real pixels and generated ones, and boundaries are where a careful look pays off.

  1. A seam along the jaw or hairline where skin tone or grain changes
  2. Ears that do not match each other, or earrings that only half exist
  3. Glasses whose frames break or bend where they cross the face
  4. Teeth rendered as one shape rather than individual ones
  5. Eyes with catchlights in different places, or none at all
  6. Skin that is smoother than the neck and hands in the same picture
  7. Hair that dissolves into the background rather than ending in strands
  8. Lighting on the face that disagrees with the shadows in the room
  9. A face slightly sharper or softer than everything at the same distance
  10. A background that is plausible everywhere and specific nowhere

Every one of these appears in ordinary photographs, and a competent operator can remove all of them. Use them to decide what to check, not to reach a conclusion.

Run a proper check instead
Side by side

Against the usual free web checker

Most free deepfake detector sites follow one pattern: take the upload, keep the picture, return a percentage with no working, and name the generator they think made it.

Capability Original or AI Typical free checker
Runs on your device, nothing transmitted Yes No Your picture goes to a server
Names the model behind the result Yes Open and published No
States its training coverage Yes 4,803 generators No
Published band thresholds Yes Five bands, line at 65 No
Reports screenshot and upscaling caveats Yes No
Separates a swap from a generated frame Partly Explained, not asserted No Collapsed into one verdict
Full-precision model Yes Quantising moved scores too far No Unstated
Works without an account Yes Partly Often after a sign-up wall
Picture of a real person stays private Yes No Uploaded and stored
Result kept only in your browser Yes No
Identifies the person in the picture No Deliberately not built No
Names the face-swap tool used No Nobody can do this reliably Partly Frequently claimed

“Typical free checker” describes the pattern shared by the free web tools we have used, not one named product. Where a competitor does better on a row, that row is wrong and we would like to be told.

See it on your own picture
Questions

The things people ask before trusting it

Using it

Can I check a deepfake for free?
Yes. There is no account and no payment. The picture is decoded and scored inside your own browser, which is why it costs nothing to offer, and the file never crosses the network. The first check takes a few seconds longer than the ones after it.
Does the picture get uploaded?
No. The file is read straight off your device, decoded by your browser and scored in the same tab. Nothing is posted to this site or a third party. On this page that matters more than most: the pictures people check here are usually of a real person who has not agreed to anything.
Why is this a separate page from the picture detector?
Because it runs a different model. The home page tool reads a frame and a tile map with its own model; this one uses a model trained across 4,803 generators. Two models with different strengths deserve two pages rather than a hidden toggle.
Why does the first check take a while?
Nothing is fetched until you actually use the tool, rather than for everyone who reads the page. What arrives is cached afterwards, so the second check starts immediately, and it arrives at full precision because a compressed version moved some scores by nearly the whole scale.
Can I check several pictures at once?
No, not here. This model is heavier than the one on the home page and reads a single frame closely, so it takes one picture at a time. For a batch of up to twenty-five, the picture detector on the home page is the right tool.

What the result means

What does the score actually measure?
It measures how closely the frame resembles the output of a generative process, on a 0 to 100 scale with the decision line at 65. It is a likelihood about the pixels, not a statement about a person, and not a claim that a specific face was swapped rather than the whole picture generated.
How do I know which part of the picture was faked?
This model returns one figure for the frame and no region map, so it cannot point at a part. The picture detector on the home page does draw a tile map, and running the same file through both is the practical answer: one gives the stronger model, the other gives the location.
Why did a swapped face only score in the sixties?
Because most of the picture is real. A swap is a small patch of generated pixels inside a genuine photograph, and the real room, clothing and light outnumber it. A middle score on such an image is an honest summary rather than a failure, which is why the middle band exists.
Does a low score prove the picture is genuine?
No. A low score is the absence of a detectable signal, not evidence that a camera was there. Screenshots, reposting and heavy compression all remove the evidence this analysis reads, so a much-shared picture can score low while being entirely fabricated.
Why was an ordinary photo of my friend flagged?
Usually retouching. Portrait modes, skin smoothing and beauty filters rearrange texture in the same direction generation does, and the model measures texture. Check the caveats shown beside the score, and if the copy you have is a screenshot, try the original file.

Limits and next steps

Does it work on video?
Yes, with limits, and on a different page. The video detector samples up to twelve frames from a clip of one minute or less and lays their scores on a timeline. It does not inspect every frame, track facial motion or measure lip sync, so a swap between sample points can be missed.
Can it tell me who is in the picture?
No, and it is built not to. Face matching is a different technology with much heavier consequences for the person photographed, and adding it would turn a privacy-preserving check into an identification service. This tool answers whether the pixels look reconstructed, and nothing about identity.
Can it name the face-swap tool that was used?
No, and no honest tool can. Attribution would require a signature unique to one product that survives compression and resizing, and none exists. Any detector naming a specific tool from a finished image is guessing, however confidently it presents the guess.
Is a result strong enough for a formal complaint?
No, not on its own. A score is a likelihood attached to pixels, and a complaint about a real person needs more than that. Use it to decide whether to pursue the matter, then gather provenance: the earliest posting, the original file, and the source photograph a swap was built on.
What should I do about a non-consensual image?
Report it to the platform first, before running anything. Most services now have a dedicated route for intimate image abuse that does not require you to prove the picture is synthetic, and organisations exist that will handle takedowns on your behalf. A detector score is at best supporting material.

Check the face you are unsure about

Drop in the picture and read what the model measured. Free, nothing uploaded, and the limits are stated beside the number rather than buried under it.

Check a picture

No sign-up. No upload. Nothing stored on our side.

Also worth reading