Google’s Nano Banana / Gemini Image family of image generation models are some of the most powerful tools for generating complex and precise AI images.

However, when created through the official Gemini interface, all generated images are stamped with an unsightly Gemini spark logo¹ in the bottom-right corner, making them unsuitable for any kind of external use.

This happens regardless of what Google AI plan you’re on. Fortunately, the watermark is pretty simple to remove. Here are a few of the best ways:

# Use reverse alpha blending

Several projects remove the watermark with reverse alpha blending. I’ve found them to be the easiest and most generally effective way of removing the watermark, subject to the following caveats:

• For some images, this method does not find the watermark correctly, creating an extra negative watermark.
• This method is consistently least effective when the watermark appears on a white background.

Pilio provides an easy-to-use, browser-based tool for doing this here, based on this source code. I’ve also had some success with this Rust crate for command-line, batch-removal of watermarks.

# Use Smart Patch in an image editor

For tough stains that just won’t yield to reverse alpha blending, I use Smart Patch in Krita, or Smart Fill from Acly’s Krita Vision Tools, depending on what works better on a given image. Other image editors definitely have similar features, and for a simple enough image you could probably achieve similar results with the clone brush.

# Use the API instead of the Gemini app

Visible watermarks are absent from images generated through the Gemini API.

However, generating images this way incurs usage-based pricing, even if you’re a Google AI subscriber. This will only put you out a few cents per image though.

The easiest way to generate images through the API is through Google AI Studio, but you could also ask a coding agent to do it for you.

# Taking it further

While I’m mostly concerned with the unsightly corner sparkle, I’d be remiss not to mention that Gemini-generated images also contain identifying metadata and Google’s invisible SynthID watermark.

Removing identifying metadata is simple – just use ExifTool.

exiftool -all= filename.jpg

SynthID removal is a much trickier problem. The most common approach at the time of writing seems to be passing an NBP generated image through img2img with an open-weight model like Stable Diffusion XL at low denoise. I haven’t had much success with this personally – in my experience it changes the image noticeably even at 5% without rendering Gemini unable to detect it. I suspect you may see better results if you use open weight models to make more substantial changes to NBP images, but I have not experimented with this very much.