SynthID Sanitizer — Disrupt Google SynthID Invisible Watermark
Heuristic detection and disruption of Google SynthID invisible watermarks in PNG and JPEG images. Rotate, rescale, and re-encode to degrade the watermark signal. Files never leave your device.
Drop a PNG or JPEG image
SynthID heuristic detection and disruption runs in your browser. Nothing is uploaded.
Intended use
- Intended for personal privacy protection: removing metadata and watermarks from files you own before sharing online. Not intended for: academic dishonesty, platform ToS circumvention, or misrepresenting content origin.
- SynthID detection is heuristic only. Google has not published the algorithm. False positives and negatives are expected.
- Disruption cannot be verified — there is no public test to confirm the watermark has been fully removed.
Google DeepMind’s SynthID is an invisible watermark embedded directly into the pixels of AI-generated images. Unlike C2PA (which attaches a cryptographically signed metadata chunk to the file), SynthID modifies pixel values in the frequency domain — the domain that encodes fine textures, edges, and noise patterns. The watermark is designed to survive screenshots, crops, resizes, and lossy compression. If you screenshot a SynthID-watermarked image and re-upload it to a different format, the watermark follows.
Google deploys SynthID across Imagen, the Gemini ecosystem, and Vertex AI image tools. Some partner platforms also produce SynthID-watermarked output through Google’s infrastructure. Because the embedding happens at the pixel level, tools that strip file-level metadata (EXIF, XMP, C2PA chunks) do not touch SynthID. You need a different approach.
This tool applies a heuristic detector that looks for statistical fingerprints consistent with SynthID embedding — frequency-domain artifacts, quantization patterns, and noise distribution characteristics that correlate with known SynthID-output distributions. It then runs a three-step disruption pipeline: a subtle rotation to misalign the frequency-domain embedding grid, a fractional rescaling to shift pixel positions, and a re-encode pass that adds quantization noise. None of these transforms is visible to the eye, but together they degrade the watermark signal that SynthID relies on.
This is fundamentally different from the C2PA Metadata Stripper. C2PA lives in a file chunk that can be surgically removed without touching any pixels. SynthID is woven into the pixels themselves. You cannot “remove” it without changing the image. The goal here is disruption — reducing the watermark signal below the threshold where Google’s (non-public) detector would return a confident positive match.
How to use
Drop a PNG or JPEG image
Drag any PNG or JPEG into the drop zone. The file is read in your browser and never leaves your device. The detector looks for heuristic indicators that suggest a SynthID watermark may be present.
Read the heuristic verdict
The tool reports whether SynthID indicators were found, along with the evidence that led to the verdict. Because Google has not published the exact SynthID algorithm, all detection is best-effort heuristic analysis.
Download the sanitized image
The disruption pipeline applies a subtle rotation, rescaling, and re-encoding pass designed to degrade the frequency-domain watermark signal without visibly affecting the image. The filename gets a `-synthid-clean` suffix.
Frequently asked
What is SynthID?
SynthID is Google DeepMind's invisible watermarking technology embedded directly into the pixels of AI-generated images. Unlike C2PA (which lives in file metadata chunks), SynthID modifies pixel values in the frequency domain — the watermark survives screenshots, crops, and format conversions. It is deployed across Google's image generators including Imagen and the Gemini ecosystem, and has been licensed to other platforms.
How does the disruption work?
The sanitizer applies a three-step pipeline: a subtle rotation (0.5–1.0 degrees) to misalign the frequency-domain grid, a fractional rescaling (99–101%) to shift pixel positions, and a re-encode pass at quality 0.92 that adds quantization noise. These transforms are individually imperceptible but collectively degrade the watermark signal that SynthID relies on.
Can SynthID be fully removed?
There is no guarantee. Google designed SynthID to be robust against common image manipulations including rotation, cropping, compression, and screenshots. Our pipeline combines multiple transforms to maximize disruption, but the watermark may persist at reduced strength. This tool is heuristic — we cannot prove removal.
Does the disruption affect image quality?
The transforms are deliberately subtle: rotation under 1 degree, scaling within 1%, and JPEG re-encode at quality 92. Most viewers will not notice a difference, but the image is no longer a pixel-exact copy of the original. If you need pixel-perfect preservation, the C2PA Metadata Stripper is a better fit — it removes only the metadata chunk without touching pixels.
Which images have SynthID?
Google applies SynthID to outputs from Imagen, Gemini image generation, and Vertex AI image tools. Some partner platforms using Google's Vertex AI may also produce SynthID-watermarked images. Images from OpenAI, Adobe, Midjourney, and Stable Diffusion typically do not carry SynthID (they use C2PA or proprietary watermarks instead).
Are my images uploaded?
No. Detection and disruption both run in your browser using JavaScript. The image bytes are read, analyzed, transformed, and rewritten entirely in-memory. No network call is made, no server sees your content, and nothing is stored between sessions.
Limitations
- Heuristic detection onlyGoogle has not released the SynthID detection algorithm or the watermark embedding key. Our detector looks for indirect statistical indicators — frequency-domain artifacts, quantization patterns, and noise distribution characteristics — that correlate with SynthID-watermarked images. False positives and false negatives are both possible.
- No guarantee of removalSynthID is designed to survive common image manipulations. Our disruption pipeline applies transforms known to degrade frequency-domain watermarks, but we cannot verify whether the watermark has been fully removed. The output should be treated as 'sanitized with best effort,' not 'certified clean.'
- PNG and JPEG onlyThe sanitizer handles PNG and JPEG inputs. WebP, AVIF, and other formats are not yet supported — convert to PNG or JPEG first if you need to process them.
- Lossy processUnlike the C2PA Stripper (which removes a metadata chunk without touching pixels), SynthID disruption requires re-encoding the image. This means the output is not a byte-for-byte copy of the input. The visual difference is negligible, but pixel values will differ.
Platform notes
- macOS
- No native macOS tool can detect or remove SynthID. Preview.app and Photos do not surface any SynthID indicators. The only reference implementation is Google's closed-source SynthID toolkit, which is not publicly available. This tool provides the closest alternative without sending files to a server.
- Windows
- Windows has no built-in SynthID awareness. File Explorer and Photos show standard image metadata but nothing related to the SynthID watermark. Any tool claiming to 'detect SynthID' without access to Google's private detection keys is heuristic-only, including this one.
- Linux
- CLI alternative: `exiftool` and `identify` (ImageMagick) can inspect pixel-level statistics but have no SynthID-specific detectors. The underlying SynthID heuristic in this tool is exposed as a pure-function module — see `src/lib/ai-cleanup/synthid.ts` in the open-source release.
- Web
- Runs entirely client-side. The file is read into a Uint8Array, heuristic analysis and disruption are performed in JavaScript, and a Blob URL is generated for the download. No upload, no server roundtrip, no telemetry.