Gemini Omni Watermark: What We Know So Far

What we actually know

Gemini Omni Flash shipped on 19 May 2026 across the Gemini app, Google Flow, and YouTube Shorts. Exactly one fact about its output watermarking is confirmed: it carries SynthID. Whether it also applies a visible corner mark, and if so what size, position, or alpha profile that mark uses, is unverified. This page states that split plainly instead of guessing at the rest, or skipping the search term entirely.

What "unverified" means in practice here

Every other model family this project supports has an entry in the watermark size catalog, built from real, checked sample output: multiple untouched samples, a consistent anchor position, and a confirmed alpha profile per size. Gemini Omni Flash does not have one yet, because building that entry requires clearing the same evidence bar as every other row in that table, not a lighter one just because the model is new. Shipping a removal claim without that evidence would be the exact doorway-page pattern this project's open scope page exists to avoid.

Why the tool fails closed here rather than guessing

If you drop an Omni Flash image into the browser eraser today, the detector runs its normal catalog lookup, finds no matching entry for Omni's dimensions and generation source, and reports nothing to remove. The file comes back unchanged. That is not a missing feature masquerading as a bug — it is the same fail-closed behavior every unsupported input gets, and it is deliberately safer than a plausible-looking geometry borrowed from a different model family. A wrong watermark geometry is not a neutral guess: applied to the wrong region of an image that never had a visible mark, it becomes damage instead of a removal.

SynthID applies regardless of what this page eventually confirms

Whatever this page learns about a visible mark, SynthID is already confirmed and out of scope for every tool on this site. This project does not detect or remove SynthID on any model family, Omni included, because SynthID is a distributed learned signal rather than a fixed pixel pattern a local detector can target. The SynthID comparison guide covers why that distinction holds regardless of which model produced the file, and it will keep holding even after Omni's visible-mark status is settled one way or the other.

What would change this page

Real Gemini Omni Flash output showing a consistent visible mark, gathered as multiple untouched samples across more than one resolution, would be enough to start the same verification process every other catalog entry went through. If that evidence instead shows Omni never renders a visible mark at all, this page will say that plainly too — an explainer that confirms "no visible mark" is just as much a finished answer as one that ships a new catalog entry. Either outcome replaces speculation with a checked fact, which is the only thing that moves this page forward.

How to help this page get finished sooner

The fastest path from explainer to tool page is real evidence, not a faster release schedule. Anyone with genuine, untouched Gemini Omni Flash output showing a consistent visible mark can contribute it through the open-source repository the same way any other catalog entry gets verified: multiple samples, a consistent anchor position, and enough resolution variety to separate a real pattern from a one-off rendering artifact tied to a single prompt or aspect ratio.