Can SynthID detect if an image was made by Gemini?

Yes. You can upload an image to the Gemini app and ask 'Was this generated by Google AI?' to check for SynthID. The feature has been used over 20 million times since launch.

Does SynthID survive editing?

Yes. SynthID is distributed across all pixels in the frequency domain and survives cropping, compression, color adjustments, and format conversion.

SynthID: What It Is, How It Works, and Why You Can't Remove It

SynthID is Google DeepMind’s invisible watermarking technology. Unlike the visible sparkle logo that Gemini stamps on generated images, SynthID operates below the threshold of human perception — embedded directly into the pixel data during image generation. It is present on every image Gemini produces, across all subscription tiers, and it cannot be removed without degrading the image. Here is what you need to know.

What SynthID Is

SynthID was developed by Google DeepMind as a tool for identifying AI-generated content. It has been applied to images, audio, video, and text produced by Google’s generative AI models. For images specifically, SynthID embeds a signal into the pixel values that is invisible to humans but detectable by SynthID’s verification algorithms.

The scale is significant. Google has reported that over 20 billion pieces of content have been watermarked with SynthID since its launch. It is not an experimental feature — it is production infrastructure running on every piece of AI-generated content that leaves Google’s systems.

How Frequency-Domain Embedding Works

Traditional visible watermarks sit on top of an image as a separate layer. SynthID does something fundamentally different: it modifies how the image is generated in the first place.

Images can be represented in two ways. The spatial domain is what you see — a grid of pixels, each with RGB color values. The frequency domain is a mathematical transformation of that same data, representing the image as a collection of patterns at different scales and orientations. Low frequencies capture broad color gradients and shapes. High frequencies capture fine details, edges, and textures.

SynthID operates in this frequency domain. During the image generation process, the model makes subtle adjustments to certain frequency components — changes so small that they are invisible when the image is converted back to pixels for display, but that form a detectable pattern when analyzed mathematically.

Think of it like an audio watermark embedded in frequencies your ears cannot distinguish from the music, but that a computer can extract. The watermark is not sitting on top of the content. It is part of the content.

Why SynthID Survives Editing

Because the watermark signal is distributed across the entire image in the frequency domain, it is remarkably resilient to common image manipulations:

Cropping: Removing part of the image removes some of the signal, but enough remains in the surviving pixels for detection.
JPEG compression: Lossy compression alters pixel values, but SynthID is designed to survive the specific kinds of information loss that JPEG introduces.
Color adjustments: Changing brightness, contrast, saturation, or applying filters modifies pixel values uniformly, but the relative frequency-domain patterns persist.
Format conversion: Converting between PNG, JPEG, WebP, and other formats does not erase the signal.
Resizing: Scaling the image up or down alters the spatial resolution, but the frequency-domain signature adapts.
Screenshots: Taking a screenshot introduces resampling and potential compression, but the signal survives.

No single pixel carries the watermark. The information is spread across millions of pixels in a way that is redundant enough to survive partial destruction. You would need to alter the image so aggressively that the resulting quality loss would defeat the purpose of having the image at all.

SynthID vs. C2PA Metadata

SynthID is sometimes confused with C2PA (Coalition for Content Provenance and Authenticity), but they work in entirely different ways.

C2PA is a metadata standard. It attaches a signed certificate to the image file that records information about how the image was created — which tool generated it, when, with what parameters. C2PA data lives in the file’s metadata headers, separate from the pixel content.

The limitation of C2PA is obvious: metadata can be stripped. If someone saves the image in a format that does not support C2PA, re-encodes it, takes a screenshot, or simply removes the metadata tags, the provenance information is gone. It is useful when the chain of custody is preserved, but fragile when images circulate freely on social media.

SynthID is pixel-level. The watermark is in the image data itself, not in the file headers. There is no metadata to strip, no certificate to remove, no tag to delete. As long as the pixels exist in a recognizable form, the watermark exists.

Google applies both C2PA metadata and SynthID to Gemini-generated images. They are complementary — C2PA provides detailed provenance when available, and SynthID provides detection capability when metadata has been lost.

The Detection Feature

Google has built SynthID detection directly into the Gemini app. If you upload an image to Gemini and ask “Was this generated by Google AI?”, the app will analyze the image for the SynthID watermark and tell you whether it detects a signal consistent with Google-generated content.

This feature has been used over 20 million times since its introduction, reflecting genuine public demand for tools that can distinguish AI-generated content from photographs. The detection returns a confidence level — not a binary yes/no — since image modifications can weaken the signal even if they do not eliminate it entirely.

It is worth noting that SynthID detection only identifies content generated by Google’s models. It cannot detect images from Midjourney, DALL-E, Stable Diffusion, or other providers. There is no universal AI image detection standard yet, though initiatives like C2PA are working toward interoperability.

Why You Cannot Remove SynthID

The reasons are mathematical and structural:

No reference signal to subtract. With the visible sparkle watermark, you know the exact overlay image and can subtract it. With SynthID, the specific encoding is unknown — Google does not publish the detection keys or the embedding algorithm. Without knowing what was added, you cannot reverse it.

The signal is everywhere. The visible sparkle affects only a small region in the bottom-right corner. SynthID is distributed across every pixel in the image. To remove it, you would need to modify every pixel, which means you are fundamentally altering the entire image.

There is no “original” to recover. The visible watermark is applied after generation — the image exists in an unwatermarked state, then the sparkle is composited on top. SynthID is part of the generation process itself. The image was never generated without the watermark. There is no pre-SynthID version to restore.

Removal attempts degrade quality. Some approaches attempt to disrupt SynthID by adding noise, re-encoding through another AI model, or applying aggressive transformations. These may weaken or destroy the watermark signal, but they also visibly degrade the image. You end up with a worse image that might or might not still carry the watermark — not exactly a win.

Honest Assessment of Bypass Claims

Do methods exist that can weaken SynthID? Yes. Heavy JPEG compression, aggressive style transfer through another model, substantial noise addition, and extreme downscaling can all degrade the watermark signal to the point where detection confidence drops below useful thresholds.

But none of these methods are free. Every one of them visibly degrades the image. The question is not “can you destroy the signal” — you can destroy any signal if you do not care about the medium it is carried in. The question is “can you destroy the signal while preserving the image,” and the answer to that, as of today, is no.

Anyone selling a tool that claims to losslessly remove SynthID is either removing only the visible sparkle (which is a different watermark entirely), or they are degrading the image more than they are admitting.

What This Means for You

If your concern is the visible sparkle logo on your Gemini images, that is a solved problem. Reverse alpha blending removes it cleanly, and tools like GeminiWM handle it in your browser with no quality loss.

If your concern is SynthID — for example, you do not want the image to be identifiable as AI-generated — you should know what you are dealing with. SynthID is a well-engineered, deeply embedded signal designed by one of the world’s leading AI research labs specifically to resist removal. The practical options for eliminating it all involve accepting visible quality loss.

GeminiWM is transparent about this distinction. It removes the visible watermark because the math allows it. It does not claim to touch SynthID because doing so honestly is not possible. Understanding the difference between these two watermarks — where they live, how they work, and what can actually be done about each one — is the foundation for making informed decisions about your AI-generated images.

Frequently Asked Questions

Can SynthID detect if an image was made by Gemini?: Yes. You can upload an image to the Gemini app and ask 'Was this generated by Google AI?' to check for SynthID. The feature has been used over 20 million times since launch.
Does SynthID survive editing?: Yes. SynthID is distributed across all pixels in the frequency domain and survives cropping, compression, color adjustments, and format conversion.