Imagen watermark remover
What the Imagen watermark remover does
Imagen, Google's image generation model, applies SynthID at generation time: a statistical pattern spread across the pixels rather than a sticker in a corner. It's built specifically to survive the ordinary edits that remove a normal watermark, cropping, resizing, re-encoding, so a simple retouch does not address it.
Our media workflow handles this with a proprietary regeneration technique. Instead of editing the original pixels, it re-renders the image: a new generation built from your file, without SynthID's original statistical signal riding along.
The problem, in plain terms
Why a normal edit doesn't touch SynthID
Most watermark removal is a pixel edit to a visible region or a strip of a metadata record, both localized changes. SynthID isn't localized. It's a sub-perceptual signal spread across the whole image, so cropping a corner or stripping a metadata field leaves it untouched, no matter how carefully the edit is done.
That's exactly why our approach is different from a retouch: it regenerates the image rather than editing around the mark.
How it works
How to handle an Imagen watermark in three steps
Regeneration runs in our media workflow, right from the panel above.
- Step 1
- Upload the original Imagen file. Starting from the actual file gives the regeneration the most to work with.
- Step 2
- The workflow regenerates the image: a new rendering built from your upload, rather than an edit of the original pixels.
- Step 3
- Download the result and compare it to the original, since regeneration is a new rendering and can shift fine detail.
After regeneration
What to check once the file comes back
Compare the regenerated image to the original at full size. Because it's a new rendering rather than a pixel-for-pixel copy, fine texture, small text, or intricate detail can come back slightly different, and that's expected, not a bug. Give both images a close look side by side before deciding the result is ready to use.
We don't promise a specific result from any third-party detector or checker on the regenerated file, and we don't promise the technique works identically on every image. Review the actual result before you rely on it.
Why it's like this
Why regeneration, not editing, is the right approach
SynthID's design goal is surviving exactly the operations that defeat other watermarking approaches: cropping, resizing, re-encoding, even a fair amount of retouching. A method that edits around the mark can't reach a signal spread through the whole image, no matter how carefully that edit is done.
Regeneration sidesteps that by not editing the original at all. It produces a new image, so there's no original pixel structure carrying the original signal to survive.
Straight answer
Does this guarantee the image passes every check?
No. We don't promise a specific verdict from any third-party detector, including Google's own SynthID check, and we don't promise an identical-looking result on every file. What we offer is a real technique, regeneration rather than a surface edit, not a guarantee attached to any single outcome.
Check the result yourself before you rely on it for anything sensitive.
Who uses it
Who works with an Imagen image, and for what reason
A few situations where regeneration is worth using.
- Developers building on Imagen
- Preparing test output for a demo without a lingering provenance mark.
- Marketing teams
- Clearing an asset before it enters a managed library with its own conventions.
- Anyone checking provenance first
- Wanting to know what a file carries before deciding what to do with it.
Practical notes
A few practical details before you upload
Our media workflow accepts JPEG, PNG and WebP, the formats an Imagen image is typically exported in, and there's no fixed limit on how many files you regenerate in a session. Each file is regenerated on its own, so nothing about one image affects the result on another.
Regeneration takes a little longer than a simple retouch, since it's re-rendering the image rather than editing a small region. That's the tradeoff for reaching a signal a normal edit can't touch.