AI Image Detector
> return_hub · module_loader · boot
Loading module…
> BOOT_SEQUENCE // ROY-HUB
HOST session.local · SEGMENT PRIVATE_NET · MEM 640K SUFFICIENT
CHECKSUM 0x9f2e · SIG VERIFIED · SESSION ACTIVE
> No clutter, no ads, no bait-n-switch—just simple tools that work at no cost
> return_hub · module_loader · boot
> return_hub · module_active
> c2pa_local · exif_local · watermark_local · sightengine_optional · no_single_oracle
Content Credentials (C2PA) and EXIF/XMP stay in this tab, and a local residual/FFT scan looks for invisible pixel watermarks (SynthID-style spread-spectrum carriers and similar). A valid trusted signature is the strongest origin signal. Missing credentials prove nothing. The optional Sightengine pass is a pixel-based hint and can be wrong on screenshots, compression, or new generators. If this tool reads the file as AI-marked, you can run a local camera-style recapture (no geometric warp) and re-run the scan. Night exteriors may be softened slightly so mid-scale generator texture does not survive. That can drop a pixel classifier score; it does not prove the picture is a photograph.
Or drop one image here. Original bytes are required — do not screenshot first if you want C2PA.
AI Image Detector reads C2PA Content Credentials and EXIF/XMP in the browser, scans pixels for a sparse spread-spectrum watermark, and can recapture the file locally to strip that mark. A valid trusted C2PA signature is evidence of origin. A missing manifest or a forensic score is not proof.
Check whether an image carries C2PA Content Credentials, read EXIF/XMP, and scan pixels for invisible spread-spectrum watermarks in your browser. If the file looks AI-marked, you can strip that watermark locally and re-run the scan. When Sightengine keys are configured, you can optionally send a copy for a pixel-based AI-generated score. A valid trusted signature is evidence of origin; a missing manifest or a forensic score is not proof.
A valid C2PA signature that names a known issuer and still matches the pixels is an origin record. A screenshot of a chat image usually has no manifest. The local scan can then only say whether it sees a sparse carrier, which is a hint.
No single detector is a proof. A valid C2PA signature from a trusted issuer (for example OpenAI, Google, Adobe, or a camera vendor) is the strongest origin signal. Missing Content Credentials are common and prove nothing. The local watermark scan and Sightengine’s pixel score are statistical hints and can fail on screenshots, heavy JPEG compression, collages, and new generators.
C2PA, EXIF, the invisible-watermark scan, and watermark stripping run in your browser on the original file bytes. Nothing is uploaded for those steps. If you enable the optional Sightengine pixel check and the server has API credentials, a copy of the image is sent to Sightengine’s API for that request only and is not stored by Roy's Tools.
C2PA is a tamper-evident provenance standard. When present and valid, it records who signed the file, whether it was generated or captured, and whether the pixels still match the signature. Re-saving, cropping in some apps, or taking a screenshot usually removes it.
If analysis reads the file as AI-marked, a strip button runs locally: camera-style recapture (optical blur, sensor mosaic, noise, chroma denoise, JPEG) plus a sparse spectral notch for lattice watermarks. The detector then runs again on the new file. This targets pixel-space marks, strips C2PA/EXIF, and can lower a Sightengine-style pixel score. It cannot guarantee a vendor’s private decoder will fail, and it does not make an AI image into a photograph.
Not for C2PA, EXIF, or the local watermark scan/strip. Pixel scoring uses SIGHTENGINE_API_USER and SIGHTENGINE_API_SECRET on the server (for example Railway). The free Sightengine plan is quota-capped. If those variables are unset, the tool still works locally and labels the pixel panel as unavailable.
Many generators never embed C2PA, and platforms strip metadata. Use the original download from the generator when you can. A screenshot of ChatGPT or Midjourney is usually unsigned and harder for both provenance and pixel models.
No. Camera make, model, and GPS are useful context but are trivial to edit. This viewer reports what the file contains, not whether those tags are truthful.