🌀 Meta Content Seal Detector
Check if an image carries Meta's open-source Videoseal / PixelSeal invisible watermark.
Drop an image here or click to upload
JPG · PNG · WEBP · TIFF · BMP. Analyzed on our server, not stored.
Technical detail: per-view decoded bits
The detector decodes 256 bits from four views of the image (original, JPEG q85, JPEG q60, 5% center crop). A real watermark survives all four; noise doesn't.
⚠️ What this tool actually detects, and what it doesn't
Detects: images watermarked using the public facebookresearch/videoseal / PixelSeal encoders. These are the open-source Content Seal image models Meta's FAIR team released.
Does NOT detect:
- Meta's production Imagine with Meta AI / Muse Image outputs. The Content Seal landing page explicitly states production uses a "custom proprietary implementation," and the production keys aren't published, so no third party can verify those images.
- Watermarks from other AI companies: Google SynthID, OpenAI DALL·E, Midjourney, Stable Diffusion, Adobe Firefly, Flux, Sora. Their detectors are either private (SynthID) or don't exist (most).
- AI-generated images without any watermark (any generator whose output has been screenshotted, recompressed, or run through a watermark scrubber).
We also run a quick metadata scan on the same upload so you get the full picture: pixel watermark AND EXIF/XMP AI markers in one result. For a deeper metadata inspection including full C2PA manifest verification, use our AI Metadata Detector.
How the detector works
Meta's Videoseal is a robust invisible watermark. It hides a 256-bit message inside an image's pixels in a way that survives JPEG compression, cropping, and mild edits, while the image looks identical to the human eye.
Our detector reads that message four times, from four slightly different views of your image (original, JPEG re-encode at q85, JPEG re-encode at q60, and a 5% center crop). If a real Videoseal watermark is present, all four decodings should agree at >85%. If nothing was embedded, the "decoded bits" are just noise and the four decodings disagree at close to chance (50%).
We report the minimum agreement across views as the confidence score. It's a zero-shot, key-free detector: it doesn't need to know what message was embedded, only whether one is consistently present.
The bigger picture
Meta published Content Seal's research models (the ones this tool uses) under an MIT license in 2024–2025. At the same time, Meta's production pipeline for its own Imagine / Muse image generator uses a proprietary variant Meta has not released. So the world can build detectors like this one, but only Meta can verify Meta's own outputs.
Meanwhile, the images Meta harvests from Facebook, Instagram, and WhatsApp to train those models routinely get their EXIF stripped by the upload pipeline. If you want to keep provenance on your photographs, our Copyright Editor and GPS Editor put you back in control before publishing.
Detector runs on facebookresearch/videoseal (MIT).
ExifReader.com is proudly brought to you by PhotoWorkout.