Simple snapshots, supposedly taken “on the fly” with an iPhone, have recently sowed doubt about the ability to detect AI in our news feeds. However, this is not just an ordinary generation on a general public tool.
At the start of 2025, photos of this young woman made the rounds on social networks, presented as the ultimate proof that creation by generative AI is now undetectable to ordinary mortals. And we can only recognize it when we see the photos: they are stunning.
The cocktail used by dal_mac on Reddit is what works best to fool people: classic, unartistic photos that could have been taken by anyone. They are not very beautiful, not very well framed and their side photo taken on the spot give an air of déjà vu to the whole series. Everyone has already taken or had their photo taken, while hiking, in front of a beautiful stone building.
The final touch is given by a style that is the most deceptive when it comes to AI-generated images: the “iPhone” photo. Our eye is used to recognizing the grain, contrasts and even defects in photos taken by Apple smartphones, as they flood social networks. Even beyond Apple, making a photo realistic “ taken by a smartphone » is an easier thing than simulating the winner of a Hasselblad prize: we know how to recognize this mainstream style, of which we have little doubt.
The same young woman at the restaurant
Still the same on a walk
The young woman exists, everything else is generated
But this affair hides another story, a little more technical, which above all shows that understanding of the possibilities of AI is not yet perfectly mastered by everyone: yes, the young woman in these photos exists. It was not generated by an AI. It is not a prompt typed with great success on Midjourney, ChatGPT or Grok, but a fine use of Flux, a generative AI tool which can be executed locally and which has the particularity of giving sufficient control to the user to obtain consistent results over several generations.
Thus, the author of the series explains having “ pushed the limits of Flux » by successively applying two models trained on the generation of this face and on the style of photos taken with an iPhone. By combining these two restricted models with a larger generation model, he achieves this result, which he continues to work on in order to make it even more realistic. How did the model used manage to create and maintain coherence on this feminine face? Well quite simply because the person in the photos exists: the sub-model “ visage » was trained with 20 real photos of a real person.
An experience on the show Unlock
This subject of the plausibility of creations generated with AI was addressed by Numerama and Frandroid in the Unlock show. For the sequence, Robin, videographer for Frandroid, had prepared real and fake images and challenged the set and the audience to detect which was which. The AI-generated images were variations of real images, modified between 40 and 60% by a model executed locally — more or less the same technique as that used for the “ the woman who doesn't exist ».
We were blown away by what we saw and we clearly didn't get the right answer every time — neither did the viewers. You can watch the sequence again in the video clip below and play the game along with us.