This free AI produces perfect people. Even naked

Aitana Lopez has 204,000 followers on Instagram (as of December 5, 2023). Her Instagram bio says she likes video games and fitness, and, discreetly hidden behind a hashtag, the entire description is just nonsense. Aitana doesn’t exist. The woman with pink hair is an illusion created by an artificial intelligence. Her story recently appeared in the media.

The text message clues for Aitana come from the Spanish agency, the ignorant one. It came about because the demand for human models decreased. In a good month, Aitana now manages to raise 10,000 euros. Almost net. Because Aitana, in complete contrast to human influencers, produces no costs, no food needs, no protein shakes, no hyaluronic creams, no union, no bathroom breaks, no sleep, no pay.

Virtual influencers are nothing new. Some of them have seven-figure followers on Instagram. The new – and surprising – is that you no longer need any skills to make Aitanas. Anyone can create Aitanas today. And it’s (almost) free.

This is made possible by the text-to-image generator Stable Diffusion. For a long time, open source software was overshadowed by Midjourney and Co. The quality was too poor, the installation too complicated and the operation too clumsy. But while several dozen programmers at Midjourney and Co. struggling with extra services, the idiosyncrasies of big companies and regulations, Stable Diffusion is run by a global army of volunteers. It is an unequal battle. Especially when it comes to portraying people, Stable Diffusion is now at least on par with the competition. When it comes to ease of use, Stable Diffusion has overtaken the competition in a sprint.

There are even more striking differences. Midjourney, DALL-E or the Bing Image Creator work via cloud computing. This means that users operate the tools via the internet. This gives providers some control. Anyone violating the guidelines will be penalized. This is not the case with Stable Diffusion. The software can be installed on a (powerful) home computer and – unlike all other popular image generators – can be independently provided with extensions. These extensions, called models, smooth out weaknesses in the AI ​​or specialize it in certain image properties. They are made by the fan community. And it’s big and works extremely hard.

For example, if you want to create images of Lego buildings, install one of the many Lego models. There are already more than ten. If you want to create particularly beautiful and realistic birds, install one of the many bird models.

With a few commands and mouse clicks, dozens of images can be generated in minutes. Much of it is still unusable. But there is always class in the audience. As you can imagine, it’s not the Lego images that drive the community in the first place. The vast majority of extensions improve the appearance of human genitals, reduce or remove clothing, and increase girth. If you want an AI guy with more than just a sock in his pants, install the “bumps” extension. If you like French maid uniforms, you will also find a suitable model. Is your creation too light for you? Muscle expansion can help. From Kambundji to Schwarzenegger, everything is possible – use the slider to create your own individual dream person. Sometimes dressed, sometimes not. AI influencer Aitana also offers paid content. Naturally.

So far so good. Fortunately, we live in a liberal society. But with the covers, the inhibitions in the field of copyright and personal rights also disappear. There are now Lora models of every little star, no matter how unknown – including those underage. Lora stands for Low Rank Adaptation, which can be used to create astonishingly real false images. The comment under the download button that the rules of use must be adhered to and respect for people must be maintained seems like a mockery given the clear preference.

The potential for misuse of stable diffusion is enormous – this is not new either. What’s new is that you don’t need a computer science degree. Creating new Lora models is also becoming easier. Hinz and Kunz get it, stubborn teens get it, even the neighborhood creep gets it. The Lora models are ‘trained’ using photos. Anyone with a public Instagram channel and a penchant for selfies provides ample material for this.

The judge will have to decide where copyright infringement of images begins. A problem with image generators is reproducing the same person. Users can remedy this by interweaving several familiar faces together: one quarter of Cage, one quarter of Chalamet, one quarter of Hemsworth. This creates more or less consistent faces. Whether the three actor models were trained with protected footage is beyond the user’s knowledge. However, research has shown that copyrighted material has been used in various models.

What will this lead to? No doubt to a tidal wave from Aitanas. The quantity and quality of fictional content will increase to such an extent that it will become increasingly difficult and, above all, more complex for consumers to distinguish fiction from reality. If that is even desired. Influencers have always played tricks with photo editing; retouched what I didn’t like. Stable Diffusion now delivers the ultimate retouching tool: it will erase thousands of real influencers. Professional agencies will specialize in fictional characters; shabby couch potatoes, legions of cheap labor will create and maintain countless fictional profiles (until they are replaced by bots). Smart influencers will take images of themselves so they can do even more chores at the same time – and finally no longer have to go to the gym every day. Of course, this tsunami will not be limited to social media. I can already see the headlines: “Japanese man marries virtual person.”

What social impact this will have remains to be seen. Of course, new problems will arise and old ones will be accentuated: the youth craze will hardly disappear, nor will the beauty craze, nor will loneliness. There will be mergers with other AI areas: digital companions will become increasingly lifelike.

At the same time, every movement causes a counter-movement. And usually things don’t happen as quickly as expected. One thing is certain: the train has arrived. He can’t be stopped anymore. And it only rolls in one direction. Stable Diffusion is just over a year old, not even in its infancy yet, but still in diapers. The current version is the worst version ever used to produce images. The virtual future is becoming even more realistic. The next big step is moving images.

You can demonize the technology. But we will have to live with it. Stable diffusion has enormous potential, precisely because of its open source nature. Even if private companies can implement regulations more quickly and efficiently, having such powerful tools alone may not be in the public interest.