Forensic image analysis
Know what made
this image.
Upload a photograph and get a calibrated verdict — real, AI-generated, or manipulated — together with the evidence behind it. Built for journalists, moderators, and investigators who need a defensible answer before they act on a photo.
5 free credits after you verify your email. No card required.
The problem
Naive AI-image detection breaks in predictable ways
Each of these is a real property of image forensics — and each one is why a specific stage exists in the pipeline below.
Recompression erases the traces forensic tools read
Techniques that read compression artifacts work by measuring the statistical residue a specific compression pass leaves behind. Every additional JPEG re-save — a download, a re-upload, a pass through a messaging app's own recompression — overwrites that residue with a new one, so a detector that leans on compression artifacts alone gets progressively less to work with the more an image has circulated.
Screenshots and re-photographs erase capture history entirely
A screenshot, or a photo of a screen or a print, does not just recompress an image — it replaces the original capture's sensor noise and compression history with an entirely new one from the re-capture device. A detector built around forensic traces from the original capture has nothing left from that capture to examine.
A small edit inside a real photo can hide from a whole-image score
A single generated or manipulated region dropped into an otherwise real photograph can be small enough, or blended well enough, that it barely moves a classifier score computed over the entire image — the same way one wrong word in a paragraph does not change the paragraph's average sentiment. Catching this case requires looking at regions of the image individually, not just the image as a whole.
A detector only knows the generators it has seen
A classifier trained against a fixed set of known image generators learns the specific statistical fingerprints those generators leave — and a genuinely new generator, or a new version of one with a materially different generation process, is not guaranteed to leave a fingerprint the classifier recognizes. Detection accuracy is a property of what a model was trained on, not a fixed law of the image itself.
The mechanism
Four stages, in order, on every scan
Every stage is listed with what it contributes and what it cannot do. The second half is the part most detection pages leave out.
Global neural analysis
Scores the whole image with a vision model trained to recognize the statistical fingerprints of known image generators. This is the primary signal and the fastest one.
What it cannot do: Cannot say where in the image a problem is, only whether one exists overall — a small localized edit inside a mostly real photo can be diluted enough that this stage alone reads it as unremarkable.
Regional analysis
Re-runs analysis over regions of the image rather than the image as a whole — this is what surfaces a small AI-generated or edited region sitting inside an otherwise real photograph, and is what produces a "manipulated" verdict as distinct from an "AI" one.
What it cannot do: A well-blended or small-enough edit can still fall between region boundaries, or produce a signal too weak to separate from ordinary image noise — this stage narrows the hardest case in the pipeline, it does not eliminate it.
Statistical forensics
Model-free, non-learned techniques reading compression artifacts, noise patterns, texture statistics, and frequency-domain traces — independent evidence that does not depend on the neural stage having seen anything like this image before.
What it cannot do: These techniques read traces that heavy recompression or a full re-capture (a screenshot, a photo of a screen) can substantially erase, leaving this stage with a weaker signal on an image that has circulated widely or been re-photographed.
Calibrated fusion
Combines the classifier score, the regional results, and the forensic features into one calibrated estimate, producing the three-class verdict — real, AI-generated, or manipulated — and a confidence tier (low, medium, or high).
What it cannot do: Its accuracy reflects the generators it was evaluated against, not a guarantee against a generator it has never seen — and its output is a probability, never a certainty. See "Evidence" below for exactly what that accuracy figure does and does not cover.
Before you act
A verdict is one input, not a final answer
truuimage is for people who need a defensible read on a photograph before they publish, moderate, or investigate. These are the cases where human judgment still matters.
Heavily re-shared images are harder to read
Every download, re-save, and re-upload — common on social platforms and messaging apps — can erase the traces an original file still carries. The more an image has circulated, the less there may be left to analyze.
Screenshots and re-photos are unreliable
A screenshot, or a photo of a screen or a print, replaces the original image with a new capture from a different device. Treat any result on a re-photographed image with less confidence than one on the original file.
Small edits inside real photos are the hardest case
A tiny AI-generated or edited region inside an otherwise real photograph may not produce a strong enough signal to flag reliably — especially when the edit is small or well blended.
Brand-new generators may score differently
Detection reflects the generators a model has seen during training. A genuinely new generator, or a new version with a different look, may not behave like the ones already in the training set.
A verdict is evidence, never proof
You get a confidence score and a tier, not a certainty. Use the result as one input to your judgment — never as the sole basis for an accusation, a takedown, or any other adverse action without independent corroboration.
Where this fits
One input into a human decision, not a replacement for it
Journalism
A reporter checking a viral image before publication runs it alongside reverse image search and source verification — a verdict and confidence tier is one more input into an editorial decision, not a replacement for the rest of that process.
Insurance claims
A claims adjuster reviewing photo evidence attached to a claim can flag images carrying a manipulation or generation signal before they are accepted as supporting documentation, and route flagged claims to closer manual review.
Marketplace and dating trust & safety
A trust-and-safety reviewer screening listing photos or profile pictures can use the confidence tier to prioritize which uploads need a human look first, rather than reviewing every submission in arrival order.
Content moderation
A moderation queue can run uploads through a scan ahead of human review and use the verdict to sort the queue — this narrows what a reviewer looks at first, it does not replace the review itself.
HR and recruitment
A recruiter checking a candidate-submitted headshot or portfolio image for signs of AI generation or manipulation before it goes into an offer package or a public team page.
Pricing
Pay for scans, not a seat
One scan costs 1 credit. Credits never expire, and a scan that fails or expires is refunded automatically — no support request needed. New accounts get 5 free credits after verifying their email, no card required.
Do credits expire?
No. Credits you buy or receive stay on your account with no expiry date.
What happens if a scan fails?
The 1-credit charge is refunded automatically — no support request needed. This covers a detector timeout or error and the rarer case of a scan getting stuck mid-analysis, which a recurring reconciliation check catches and refunds on its own once the scan has clearly stalled.
Questions
Before you sign up
What is the difference between an "AI" verdict and a "manipulated" verdict?
"AI" means the image reads as wholly synthetic. "Manipulated" means part of the image looks real and part looks generated or edited — a localized change inside an otherwise real photograph.
Is my image stored after I scan it?
In the ordinary case, yes — the original file you upload is deleted within seconds of the scan finishing, whether it completes, fails, or expires. Two honest edge cases: a stuck scan can take longer (up to roughly ten minutes, while an automated job expires it), and if that job's credit-refund step keeps failing, the original is retained until that failure is fixed. What is retained either way is a small, resized copy — capped at 512px on the long edge, re-encoded, with EXIF, GPS, and other embedded metadata stripped — plus the verdict and forensic report, so your scan history has something to show you without keeping the original around.
See the full retention breakdownIs a verdict admissible in court, or otherwise conclusive?
No. A verdict is a calibrated, probabilistic estimate — a confidence score and a tier, not a binary certainty — and it is one input to a human judgment, never a determination on its own. It should not be the sole basis for an accusation against a person, a takedown decision, or any other adverse action without independent corroboration.
What image formats and sizes are supported?
JPEG, PNG, and WebP. The web console accepts uploads up to 15 MB, because your browser sends those bytes straight to storage rather than through our servers. The API is capped lower, at 4 MB, because an API request's bytes travel through a serverless function with its own platform-level body-size limit — the two limits differ for that reason, not by oversight.
Is there an API?
Yes — a REST API with hashed API keys is part of the product, not a bolt-on: authentication, the detect and poll endpoints, error codes, and working examples in curl, JavaScript, and Python.
Read the API docsHow do I delete my account and my data?
From your account settings, type your email address to confirm and submit — deletion is processed immediately, and you are signed out the moment you confirm. Your uploaded images, thumbnails, name, password, two-factor secret, API keys, and Stripe customer record are deleted; your email address is freed for reuse. Your scan history is retained with the filename and image fingerprint removed, and your purchases, credit ledger, and balance are retained as anonymised financial records — linked to your account id, not to you — because we keep those for accounting purposes.
See the full retention breakdownGet started
Create an account and run your first scan
New accounts get 5 free credits after verifying your email — no card required. One scan costs 1 credit.
Starter
$9.00
50 credits
Pro
$39.00
250 credits
Studio
$129.00
1,000 credits