The job
Where the SynthID remover sits in a process
The upload panel sends a file into our media workflow, where detection and the edit run together and a regenerated file comes back. It is the third step in a sequence, not the first, and treating it as the first is how people end up processing files that were never marked.
The two steps before it are identification and verification. Identification means working out which of four SynthID types you are dealing with, because Google applies the technology differently to images, video, audio and text. Verification means confirming a mark exists. Both cost nothing and both are skipped constantly.
Step one
Identify which SynthID you are dealing with
Four types, four mechanisms, and only some of them are a file you can upload anywhere.
- Image
- An invisible watermark in the image data. Google says it does not change image quality and is designed to survive cropping, filters and lossy compression. Uploadable.
- Video
- The same approach applied per segment, with frame rate changes named among the modifications it is designed to withstand. Uploadable.
- Audio
- Applied to Google's Lyria music output and the NotebookLM podcast feature, staying inaudible and surviving noise, MP3 compression and speed changes.
- Text
- Not in the file. Google says it adjusts the probability scores behind token selection, so the mark is the pattern of word choices. Nothing to upload and nothing to delete.
Step two
Verify the file is marked before you process it
Google publishes the check and it is free: upload the image, video or audio clip into Gemini and ask whether it was created or altered by Google AI. Gemini will check for a SynthID watermark. Google also runs a SynthID Detector portal, though access is aimed at journalists, media organisations and researchers rather than open to everyone.
Build this into the process as a gate. A significant share of files people bring to a tool like this were never marked, either because they came from a non-Google generator or because they were edited through a pipeline that discarded the original data. Verifying takes a minute and it saves you from evaluating a regenerated file against a problem that did not exist.
Step three
How to remove a SynthID watermark in three steps
Once identification and verification are done, the run is short.
- Step 1
- Keep the original file. There is no undo inside the workflow and the output is a separate rendering, so the source is your only comparison.
- Step 2
- Upload one file. Detection and the edit run together and a regenerated version comes back rather than your original with something scraped off it.
- Step 3
- Review at full size, weighting attention toward busy regions and fine texture, which is where a re-rendering differs first.
The branch that leaves this workflow
Watermarked text is not a file problem
If your SynthID question is about a document, this workflow ends here and no upload will help. Google's developer documentation describes SynthID for text as a logits processor applied after Top-K and Top-P, using a pseudorandom function to adjust the model's own logits as tokens are chosen. The watermark is a statistical property of the word choices, so there is no byte to strip.
Two of Google's documented limitations are worth carrying into any process that touches generated text. The watermark is less effective on factual responses, because there is less room to alter generation without hurting accuracy, and detector confidence scores can be greatly reduced when the text is thoroughly rewritten or translated into another language. Those are Google's statements about its own detector, and repeating them is not the same as us promising an outcome.
Our method, stated
Regeneration, and what it means for the review step
The workflow regenerates the file. It re-renders your image or clip rather than peeling a marker off the original, so what you download is a new rendering of the same content. That is why review is a step in this workflow rather than an optional extra: you are evaluating a new artefact, not checking that a correction landed cleanly.
It also sets what we will not say. No guaranteed result on any given file, and no claim about what any detector returns afterwards, Google's own included. If a downstream process depends on a specific detector verdict, that dependency is on a third party's system and no tool here resolves it.
The gate before all the others
Ownership, labelling rules and intent
Run this file through your own policy before you run it through anything else. Work on files you own or are entitled to edit. Some jurisdictions impose AI content labelling requirements, and those attach to what you do with the content rather than to the bytes in it.
The uses that stand up are ordinary: archiving your own generated work, research and media forensics, and getting your own files through tooling that mishandles marked input. Presenting generated work as somebody's hand-made work is not a technical problem and nothing in this workflow addresses it.
Where it fits
Who builds a SynthID step into a process
Teams with libraries of their own generated media, where the mark is an artefact of the generator rather than something they want carried forward. Researchers and journalists examining detection behaviour, which is the audience Google is piloting its own detector portal with. Production teams whose downstream tools behave oddly on marked files.
And a large group who stop after step one. They came to find out what SynthID is and which kind they have, and for them the identification table is the whole answer.