Google’s Gemini 2.0 Flash AI Model Is Surprisingly Good at Removing Watermarks from Images

Some of the latest “experimental” features in Google’s Gemini 2.0 Flash model have raised serious concerns among many users. In particular, the model has shown “very good” results in removing watermarks from images.

Image source: Google

Gemini 2.0 Flash can not only generate images based on a text request, but also edit them according to the user’s instructions – examples are published on Reddit. The other day, it was discovered that the model can remove watermarks with high accuracy. This behavior is radically different from the GPT-4o model from OpenAI, which rejects requests to remove watermarks.

There have been tools like Watermark Remover.io that can remove watermarks from stock photos, and a Google research team created a similar algorithm in 2017 to highlight the need for stronger image copyright protection.

Gemini 2.0 Flash seemed to do the best job of removing complex watermarks, like Getty Images stamps, and was able to restore the image well. Of course, after removing the watermark, the model adds a SynthID tag, effectively replacing the copyright label with “edited by AI.” But AI tags are fairly easy to remove with other AI, as Samsung’s Object Erase tool has previously demonstrated.

According to user reports, the lightweight Gemini 2.0 Flash can also add recognizable images of real people to photos, something the full Gemini model cannot do.

At the moment, the described capabilities are only available to developers via AI Studio — the publicly available Gemini bot refuses to remove the security signs, as shown in the screenshot above. Google did not respond to inquiries about the availability of protection against possible abuse after the model is released to the public.