Google Lets Users Turn Off Visible Watermarks on Nano Banana, Omni, and Lyria Outputs

Google will let users remove the visible watermark from AI-generated images, videos, and songs, the company announced Friday. The change applies to content made with the Nano Banana, Omni, and Lyria models, and does not affect Google's invisible SynthID watermark or C2PA metadata standards.

Josh Woodward, Google's VP for Gemini, confirmed the toggle in a post on X. The setting will first appear in Gemini and Google's video editor Flow, with support for Search "coming soon," according to Woodward. Once live, users can navigate to Settings > Media Watermark to switch the visible mark on or off. Google says the feature is rolling out over the coming days.

According to Woodward, the move balances "creative control and safety": visible watermarks become optional, but invisible SynthID watermarks and C2PA metadata continue to run in the background for transparency purposes. Users can still check whether an image was AI-generated using Gemini or Search, he said.

Google is also open-sourcing a new library called Credentio, which the company says lets developers embed a local validation mechanism directly into their own apps to check content provenance without relying on Google's infrastructure.

The announcement lands shortly after Anthropic's watermarking move for Claude-generated text and files, implemented to comply with EU regulations — a decision that drew criticism from some users and developers. Google's approach differs by keeping the underlying detection layer (SynthID, C2PA) mandatory while making only the visible, on-image marker optional.

What this means

Visible watermarks have long been a friction point for professional and creative users who need clean output for client work, editing pipelines, or publishing — a watermark stamped across an image or video often makes it unusable without extra editing steps. By decoupling the visible marker from the underlying provenance system, Google is betting it can satisfy both camps: creators get unmarked output, while regulators, platforms, and fact-checkers can still rely on SynthID and C2PA metadata to identify AI content programmatically.

The practical effect is that visible watermarks — which were never a strong technical safeguard to begin with, since they can be cropped or edited out — lose their remaining purpose as a casual deterrent. The real burden of detection now shifts entirely to invisible signals and metadata standards, which depend on downstream platforms actually checking for them. Google open-sourcing Credentio suggests it wants that checking infrastructure to spread beyond its own products, but adoption by third-party apps and platforms is not guaranteed. As more labs face similar EU and platform-level disclosure requirements, expect this split between visible and invisible watermarking to become a common industry pattern.