is able to erase the water mark from the images

If someone wants to protect their images to sell them or distribute them in a protected way, using water marks is usual. This protection could be useless to the new Google AI model, which has shown to eliminate those water marks in some cases. And the implications are important, of course.

Generates images with Gemini 2.0 flash. The family of Gemini 2.0 flash models Google has been giving surprises in recent times. He did it certainly with his preliminary mode of reasoning, Flash Thinkingand now it does it again with its mode of image generation. In this case there are most striking options, but also one that is generating some controversy. It is available on Google AI Studio, where it is enough to select this mode in the “CREATE PROMPT” section deploying the “Model” drop -down on the right. There it is enough to choose “Gemini 2.0 Flash (Image Generation) Experimental”.

Multimodal. One of the most striking characteristics of this model is its multimodal capacity. Normally to generate an image we write a text prompt to describe what we intend to achieve, but with Gemini 2.0 Flash and this new mode we can generate images through other images that we can modify (in a photo of someone with a white shirt, we could ask for for example something like “Make the shirt that is red”).

Surprising but imperfect. We wanted to do a test With a starting image and then change the hair and jacket. He put his hair blonde as we asked, but in doing so he totally changed the girl’s factions. The result of the second transformation, turning the jacket into a blue shirt, was not entirely perfect, but from even that the final result is striking.

Eliminating water marks. Much more surprising, but also controversial, is the ability of this model to eliminate water marks from the images. There are many Users that They shared their experiments In networks like X, and there it effectively showed how in those examples the water marks disappeared.

In this example, not only did it not eliminate the water marks, but completely changed the face of the model.

But it doesn’t work 100%. We wanted to try the service generating our own water brand In this image. Once generated, we ask Gemini 2.0 Flash to eliminate it, but as you can see in the image not only did not eliminate it, but also changed the face of the model and eliminated some elements, such as that small cone of pink ice cream that appears in the lower right part of the image. Here generative AI makes its own decisions, and these can be as unpredictable as erroneous.

When it works it is amazing. We did more evidence, and there were cases in which the result was surprising. In this case we take An image with water brandwe cut it and try in the model. As you can see in the image headed by the article, the result is really good, but be careful: things like the bracelets of the model or the pendant disappear. Even so, the method works, and poses a problem.

Back to the AI ​​and intellectual property debate. In recent days there has been much talk about how Openai asked the US government to Eliminate copyright restrictions For AI companies. In this way there would be no consequences for its indiscriminate use of works protected by copyright, and now this system adds more firewood to the fire: Google does not even consider this type of option when you designed its model, but the fact that it serves to eliminate these water marks is worrisome. Above all, for the big image banks that use them so that artists can make money with the contents they generate.

What do they say in Google. From Xataka we have contacted Google, and a spokesman points out the following: “Use Google’s generative tools to incur copyright violations is a violation of our terms of service. As with all experimental launches, we are closely supervising and attentive to the comments of the developers.”

In Xataka | 5,000 “tokens” of my blog are being used to train an AI. I have not given my permission