When Google generates an image through Gemini, it doesn't just create pixels. It embeds an invisible watermark — called SynthID — directly into the image's frequency spectrum. This watermark is designed to survive everything that strips traditional metadata: screenshots, resaving, compression, sharing through messaging apps, even cropping and resizing.
Today, we're adding watermark forensics to every Sightova scan.
The Problem with Metadata-Based Provenance
The digital provenance ecosystem has long relied on metadata. Standards like C2PA embed signed credentials into file headers. EXIF data records camera information. XMP stores processing history.
The fundamental weakness is obvious: metadata lives in the file container, not in the image itself. Save the image to a different format and the metadata is gone. Send it through Telegram, WhatsApp, or Instagram — stripped. Take a screenshot — completely lost.
This means metadata-based provenance only works in controlled chains of custody. The moment an image leaves that chain — which happens almost immediately on the open internet — the provenance data disappears.
How Invisible Watermarks Work
SynthID and similar systems take a fundamentally different approach. Instead of storing information alongside the image, they embed it inside the pixels by manipulating the image's frequency domain.
Every digital image can be decomposed into frequency components using a Fourier transform. Low frequencies represent broad color gradients. High frequencies represent sharp edges and fine detail. The key insight behind spectral watermarking is that you can add carefully controlled energy at specific frequency positions — called carrier frequencies — without visibly altering the image.
The watermark becomes part of the image itself. You can't strip it by changing file formats, because it's not in the file format. You can't remove it by clearing metadata, because it's not metadata. It persists through compression, rescaling, and format conversion because the underlying pixel relationships are preserved.
What We Built
Our watermark forensics module performs spectral analysis on every image scanned through Sightova. The process works in three stages:
Noise Extraction — We separate the image's content from its noise residual using denoising filters. In a watermarked image, the noise residual contains the watermark signal.
Carrier Frequency Analysis — We compute the Fast Fourier Transform of the noise residual and measure energy concentration at known SynthID carrier positions. Research has identified 28 specific frequency positions where SynthID embeds its signal. If these positions show elevated energy compared to the surrounding spectrum, it indicates a watermark is present.
Phase Coherence Measurement — SynthID carrier frequencies come in symmetric pairs with specific phase relationships. We measure whether these phase relationships are consistent with the known SynthID pattern. High phase coherence at carrier pairs is a strong indicator of an embedded watermark.
The analysis runs across all three color channels, with the green channel weighted highest — SynthID embeds its strongest signal there.
What You See in Results
When watermark forensics detects a SynthID watermark, you'll see a new section in your scan results labeled Watermark Forensics. It reports:
- Detection status — whether a known invisible watermark was found
- Confidence score — based on combined carrier energy and phase coherence
- Provider identification — currently Google Gemini for SynthID
- Carrier Energy score — how much signal is concentrated at known watermark frequencies
- Phase Coherence score — how consistent the spectral phase relationships are
- Per-channel breakdown — energy ratios for Red, Green, and Blue channels independently
This information appears alongside your existing detection results — authenticity probability, generator identification, heatmap, and content safety scores. It's another independent signal in the forensic picture.
Why This Matters for Detection
Watermark forensics and AI detection are complementary, not redundant. They answer different questions through different methods:
AI detection (our ViT models) identifies synthetic images by analyzing visual patterns and statistical artifacts left by the generation process. It works on any AI-generated image regardless of whether it was watermarked.
Watermark forensics (spectral analysis) identifies images that a specific provider has marked. It confirms origin — this image passed through Google's generation pipeline — with an independent method.
When both signals agree — our detection model says AI-generated, and watermark forensics finds a SynthID watermark — you have two completely independent confirmations from fundamentally different analytical approaches. That's a level of certainty neither method provides alone.
The Limitations We're Transparent About
Our current watermark forensics implementation is based on publicly available research into SynthID's spectral structure. Without access to Google's proprietary encoder or a reference codebook built from thousands of confirmed Gemini images, our detection is heuristic rather than exact. We're transparent about the confidence scores — they reflect what spectral analysis alone can determine.
As the research community expands watermark analysis to other providers — Meta, OpenAI, Stability AI all have their own watermarking systems — we'll expand our forensics accordingly.
The Bigger Picture
The AI provenance landscape is bifurcating into two approaches: metadata standards (C2PA, IPTC) that require cooperation from every link in the content chain, and invisible watermarks that persist regardless of how the image is shared.
Both have roles to play. But for detection purposes — for answering the question "was this image generated by AI?" — invisible watermarks offer something metadata cannot: persistence. You can't accidentally or intentionally strip a spectral watermark the way you can strip an EXIF tag.
By integrating watermark forensics alongside our detection models, metadata analysis, and pixel-level heatmaps, Sightova provides the most comprehensive forensic picture available for any image you need to verify.
Try the new watermark forensics today — it's included in every scan at sightova.com.