Remove Watermarks from Images with Google AI

Gemini AI Model and Watermark Removal

Social media users have recently highlighted a contentious application of Google's new Gemini AI model: the ability to eliminate watermarks from images. This includes images originating from sources like Getty Images and other prominent stock media providers.

Expanded Access and Capabilities

Google broadened access to the image generation functionality of its Gemini 2.0 Flash model last week. This feature allows for both native image creation and editing. The capability is demonstrably powerful, yet appears to lack sufficient safeguards.

Gemini 2.0 Flash will generate images featuring celebrities and characters protected by copyright without objection. Furthermore, as previously indicated, it can remove watermarks from pre-existing photographs.

Filling Gaps and Skill Level

Users on platforms such as X and Reddit have observed that Gemini 2.0 Flash doesn't merely remove watermarks. It also attempts to reconstruct the areas where the watermark was previously located.

While other AI-driven tools offer similar functionality, Gemini 2.0 Flash appears to excel in this area, and is currently available without cost.

Limitations and Current Status

It’s important to note that the image generation feature of Gemini 2.0 Flash is currently designated as “experimental” and “not for production use.” Access is limited to Google’s developer tools, such as AI Studio.

The model isn’t flawless in watermark removal. Gemini 2.0 Flash encounters difficulties with certain translucent watermarks and those covering substantial portions of an image.

Copyright Concerns and Restrictions

Despite these limitations, some copyright owners are likely to object to the unrestricted use of Gemini 2.0 Flash. Other models, including Anthropic’s Claude 3.7 Sonnet and OpenAI’s GPT-4o, explicitly decline to remove watermarks.

Claude, for example, identifies watermark removal as both “unethical and potentially illegal.”

Under U.S. copyright law, removing a watermark without the permission of the copyright holder is generally prohibited, with limited exceptions.

Google’s Response

Google did not immediately provide a response to a request for comment received outside of regular business hours.

Update (March 17, 1:48 p.m. Pacific): A Google spokesperson issued the following statement:

“Violating our terms of service by using Google’s generative AI tools for copyright infringement is prohibited. We are closely monitoring this experimental release and actively seeking developer feedback.”