Search for clues in every picture.
ImageDETECTIVE is a tool that checks images for signs of AI generation and manipulation. Start an investigation, examine the evidence, and understand the findings.
Steps to use
Three steps. A clearer picture.Choose a photo from your device, drag it into the box, and start your investigation.
ImageDETECTIVE checks the whole image, signs of AI editing, every face it finds, and the file's metadata.
Read the assessment and compare the evidence view with the original image.
Examine an Image
A closer look starts hereYour image is checked on this app's server and kept in memory for at most 30 minutes; it is never saved to disk. If it is not clearly authentic, a copy at most 1,024 pixels long is also sent to Meta's API (standard mode; Meta doesn't train on it) so the investigator can name and outline the evidence.
The Investigative Report
An assessment backed by evidence. Not just a label.
Your investigation starts with an image.
Upload a photo above to begin.
The colours are the whole-image check's own signals for each part of the picture: model signals, not proof of authenticity and not exact edit boundaries. The AI-edit check judges the picture as a whole, so it has no map. Numbered boxes mark faces: red looks AI-made or swapped, orange shows some signs, green shows no clear signs, grey is too small to check.
Investigation details
Try to fool it.
Describe a picture. Meta's Muse Image makes it, and ImageDETECTIVE investigates it right away. Then crop and re-save it, like a forwarded copy that has lost its AI label, and investigate again.
Text prompts only: this page never edits uploaded photos, and it doesn't make pictures of real or famous people.
About ImageDETECTIVE
ImageDETECTIVE is an image-authenticity investigation tool. It combines visual detectors, available file metadata and optional AI-assisted explanations to help people assess signs of AI generation or manipulation.
An investigation provides evidence, not proof. "Likely authentic" does not establish who took an image, where it was taken or whether its caption is true.
How the investigation works
One upload runs every check on this app's server. The verdict comes only from ImageDETECTIVE's own detectors. For pictures that are not clearly authentic, Meta's investigator then points to the most suspicious object and writes a short explanation from the detectors' numbers; it never decides the verdict.
| Component | Role | Limitation |
|---|---|---|
| DINOv3 (Meta) with a calibrated head | Whole-image check; its patch signals make the evidence view | A coarse map (one cell per 16 px of the checked copy), not an edit boundary |
| Perception Encoder (Meta), fine-tuned | AI-edit check: traces an AI image model leaves when it re-renders a photo; a second head for pictures rendered by AI image models | Judges the picture as a whole; small local edits are caught less often |
| SAM 3 (Meta) + face checks | Finds every face; each is checked for being AI-made and, with a separately fine-tuned Perception Encoder, for being swapped | Faces under 64 px are not checked; face swaps are only partly covered, less in heavily compressed copies |
| File metadata | Reads AI labels written into the file (IPTC "trained algorithmic media", C2PA presence) | Forwarded copies and screenshots lose them, so a missing label proves nothing |
| Muse Spark + SAM 3.1 (Meta's API) | Investigator: names and outlines the most suspicious object; writes a short explanation from the numbers | Optional; may be unavailable. The outline shows where to look, not a proven edit boundary |
| Muse Image (Meta's API) | Makes the pictures on the "Try to fool it" page | Text prompts only |
How the verdict is set
The thresholds are set on development photos so that at most 5% of real photos get an AI verdict. When the checks don't point clearly either way, the answer is "Uncertain" instead of a guess. The percentage shown is the whole-image check's calibrated AI evidence; the other checks are shown as findings, not as probabilities.
How to read the evidence colours
- Red: parts of the picture whose signals point to AI generation or manipulation.
- Pale green: parts whose signals point to a real photo.
- Dark blue: weak or mixed signals.
- Numbered boxes mark faces (red: looks AI-made or swapped; orange: some signs; green: no clear signs; grey: too small to check). A yellow outline, when shown, is the object Meta's investigator pointed to.
Known limitations
- It can't tell when or where a photo was taken: a real photo forwarded with a false story still looks real.
- Classic manual edits (old-style Photoshop cut-and-paste) are not covered; face swaps and small AI edits only partly.
- Some newer image generators are caught less often once a picture has lost its file label (for example in a screenshot).
- Built for photos: drawings, documents, chat screenshots and videos are out of scope.
What happens to your upload
- The image is checked on this app's server and kept in memory for at most 30 minutes (for the investigator step); it is never saved to disk.
- For images that are not clearly authentic, a copy at most 1,024 pixels long is sent to Meta's Model API in standard mode (Meta doesn't train on it). The explanation is written from numbers only.
- "Try to fool it" sends only your text prompt to Meta's API.
Source code
Code, measured results and dataset credits: github.com/phantom007xy/ImageDETECTIVE ↗. Built with DINOv3, Perception Encoder, SAM 3, SAM 3.1, Muse Image and Muse Spark (Meta).