Google’s AI model has gotten pretty good at spoofing phone photos

I’m starting to understand where Google’s visual AI model gets its name, because after playing with it for a few days, this is how I’ll sum it up: bananas. The images it generates are so realistic that they are bananas. I feel like I’m going bananas after staring at them for so long. If I had to pinpoint one reason why the Nano Banana Pro’s photos look more realistic than the AI ​​that preceded it, it’s this: They look like photos taken with a phone’s camera.

Sure, the stories are there if you look for them. Take the photo at the top of this article of the (not real!) couple on the city sidewalk. The street light in the background doesn’t look quite right to me, and some of the building facades – especially in the background – look a bit weird and blocky. But if I scroll through this photo on social media? It’s impossible for me to register it as artificial intelligence. The subjects look realistic, but I think the fact that the image doesn’t look too perfect is what appeals to it.

Bright, flat exposure, generous depth of field, and slightly crunchy detail: it all screams phone camera to me. Ben Sandowski, co-founder of the popular iPhone camera app Halide, agrees. In the AI-generated image of the ferry boat above, he points out “the aggressive image sharpening you encounter in smartphone photos. It’s an optical trick that helps the image ‘pop’.” Is there another distinguishing feature of photos taken with a phone? Noise. “Most AI-generated images look very clean. The texture in these images looks like they were taken from a small smartphone sensor.”

So where does Google’s AI get its concepts about phone photos? Google Photos may seem like an obvious place to be — and very problematic — but Elijah Lawal, director of global communications for Gemini, says: “For Nano Banana, we don’t use Google Photos.” He also told me that the Nano Banana Pro isn’t specifically geared toward producing a phone camera look. “One huge improvement is that it can connect to Google Search,” he says. If you ask it to create an infographic about today’s weather, it can look up the temperature — previously, you would need to include more of this information in your prompt.

According to Lawal, this is limited to text search and not image search. But to be able to Go and get Real world information itself may be a key element here. The Nano Banana Pro is particularly good at adding things to photos that make sense in that context — even if you don’t specifically ask for them. Historical elements can be added, such as: Period-appropriate clothing and cars Without being explicitly asked to do so. He even added a watermark to Northwest Multiple Listing Service When I asked her to create a fake Zillow listing for a fake house in Seattle. It’s much better to understand the task and add those little details without being asked.

I asked Gemini for a Zillow listing for a Craftsman-style home with white paint and black trim in West Seattle. It returned a text-only list describing the place, but with another prompt, I used Nano Banana Pro to create an image that matched the description. I didn’t specifically ask for this, but in the photo there is a Copyright 2023, which is very funny, and a watermark like the one you find on basically every real estate photo you find in the greater Seattle area. Interestingly, it’s not the current logo — it’s the previous version, the same logo that’s in every photo of the house I bought in 2018.

I asked Google where Nano Banana could have come up with this, and DeepMind product manager Naina Raisinghani suggested it was just a hallucination, offering this statement: “Nano Banana Pro offers significant upgrades to character consistency, image generation, and search-based accuracy. While this is our most accurate image model to date, AI hallucinations can occur. If an image isn’t quite right, we encourage you to try again, as a subsequent attempt often yields a more accurate result.” In keeping with your intention.” The thing is, adding the watermark to the real estate listing service the model seems to work exactly as intended.

Watermark or not, I think the small print on the “for sale” sign might indicate that this is an AI, or maybe the potted plants on the front porch look a little too perfect, but honestly? I find it hard to believe that this house isn’t real, even though deep down I know it isn’t. I wouldn’t think about it again if I found it on a real estate site, and the watermark would certainly help sell it as original. If AI gets good at imitating things that indicate a photo is real, then guys: we’re cooked.

This is what’s most worrying to me: it’s getting harder to detect what the AI ​​is saying, and the Nano Banana is getting better at simulating the little details that make an image look real. We gave her some vague prompts to portray A edge A reporter covering a live event; Added details like microphone with edge The logo is in the reporter’s hand and the crown is at the bottom of the screen. There are no misspellings or strange-looking characters. No hand with six fingers. Nothing that would clearly mark it out as artificial intelligence and a lot of little details to sell it as the real deal.

A year ago, or even a few months ago, I had a feeling that there was a day coming in the future, a day when it would be unwise to believe any photo or video I saw online from an unfamiliar source unless proven otherwise. This exercise has convinced me that that day is not in the future; He’s here now. Set your AI radar appropriately, and don’t be surprised if it leads you to a few bananas.