Images created by Midjourney: 5 tips for avoiding the trap

The images are close to perfection, but the flaws remain. The fifth version of Midjourney, available since March 15, allowed Internet users to take a series of frames that made it into the news. In less than two weeks, they showed Donald Trump in prison, Emmanuel Macron in front of the police, Pope Francis in a down jacket, or Vladimir Putin in front of a court in The Hague. All this with such realism that deceived more than one. However, there were a few clues hidden in these creations. Here we give you some tips to avoid falling into the trap.

1. Check the context

As with every post that appears on social media, the first step is to go back to the source. So a reverse image search via Google can lead to the first post location, specifically on the online Reddit forum. It is in one such discussion that over 150,000 members share their creations daily on Midjourney. On social networks, a description or comments that accompany the photo will also come in handy. Internet users tend to indicate that they have used the Midjourney tool, and the #IA (for artificial intelligence) hashtag often accompanies these digital works. For example, journalist for 50 photos false arrest of Donald Trump explained on Twitter that he was having fun “create” these pictures.

2. Pay attention to aesthetics

If no context is provided, the image’s overall first impression of “too good to be true” is a good indicator. Originally designed for artistic purposes, Midjourney produces works with unrealistic grain because they are too perfect. This plastic aspect can be found on faces, hair, and also on clothes, especially when they are solid colors, as in this orange jumpsuit from Donald Trump. In addition to the unusual texture, perspective and proportion errors create an overall surreal impression. It’s the same with shadows, often very aesthetic, but not logical.

3. Watch your hands

The second element worth paying attention to is the hands. If Midjourney has made huge strides in this area, this body part continues to be one of the weak points of this software. So, in the image below, Emmanuel Macron has six fingers on his right hand. The AI ​​drew five of them, forgetting to take into account the thumb, which should be hidden in a clenched fist. The tool also ignores how the fingers intertwine or how they hold an object.

More broadly, it is critical to check the consistency of the entire body whose symmetry the instrument is trying to reproduce. This is especially true for the ears and eyes. The lack of reflection in the latter is also a clue to take into account.

4. Scan all objects

Apart from the body, all objects must be carefully examined. Because if Midjourney knows how to replicate their aesthetic, he doesn’t understand their usefulness. Result: it offers completely bizarre shapes. This shot of Pope Francis in a down jacket is a perfect example. The ultra-realistic creation – and now viral – has several anomalies. The supreme pontiff wears three rings, which is not normal, and the cross he wears around his neck forms a gray zigzag. Nothing to do with the image of the crucified Jesus. Finally, what he is holding on the left is completely strange and does not resemble any existing object.

Other objects may also be absurd in their context. For example, on the road, artificial intelligence runs the risk of making the wrong sign colors, displaying a left arrow when making a right turn, or even placing a safety mirror in a completely out of place. The photo below is an illustration.

5. Check texts

Like objects, Midjourney ignores how writing works. If the program is going to copy a font or aesthetic, the letters won’t matter. Thus, the world famous McDonald’s brand becomes Minonade. The letters “POLICE”, stamped on the vests of the forces of order, turn into “KLIOT”. In addition to well-known names, the letters are sometimes distorted, the texts do not make sense, and the names of the signs are surreal.

Finally, the last solution in case of doubt lies in the Mayachitra tool, the GAN detector or “generative adversarial network”, named after these networks, which compete with each other to create the most realistic image. This is the technology that powers Midjourney. The results of this site are not yet reliable enough to be conclusive, but they provide hints.

Therefore, vigilance is required before introducing a reliable verification tool or watermark to identify AI-generated images. So far, only a critical approach to each image can avoid pitfalls. Check the signature, enlarge the image, scan the background, pay attention to the details. In an age of disinformation, taking a position of doubt is more necessary than ever.

Do you want to ask us questions or provide information that you do not believe is reliable? Feel free to write to us at lesverifié[email protected]. You can also find us on Twitter: our team is there on @verif_TF1LCI.