Lock, Stock and Two Smoking Bairrels
Rédigé par Amélie Guillaumot & Romain Pialat
-
08 September 2025To round off this series of experiments in style, let's dive into the topic of persona creation, asking ourselves how AI could eventually invent someone.
In this article, we try our hand at using other generative AIs besides ChatGPT, moving away from text generation to image creation. Welcome to the era of deep fakes !
Amelia
Published on 14 April 2023
« Amelia is 25 years old. She is a passionate scientist who works in a biology research team. She is particularly interested in plants and their role in the environment. She is an intelligent and determined young woman who does not shy away from challenges. She is curious and always eager to learn more about topics that interest her. Outside of work, she enjoys spending time in nature and participating in various outdoor sports. She has dark brown hair and green eyes. She is about 5'5“ tall and fairly slim. She often wears comfortable, practical clothing to work, such as jeans and sneakers ».
A photograph portrait of Amelia taken in the early 1970's, grainy, taken on a polaroid camera, realistic, hyperrealistic.
Amelia does not exist. Her story, her profession, and her personality were defined by ChatGPT. Based on this single rule : « Create a fictional character for me », the AI generated a description of about ten lines.
The text generated by the AI was then transformed into a prompt for Stable Diffusion to produce a set of images depicting Amelia's portrait.
The process is not as fast as when using existing content or the name of a celebrity. How can consistency be achieved between the different portraits of Amelia ? Not to mention ghost members, blurry eyes, or environments that don't match the context... Amelia's profile was built through several « funnel » phases : from selective sorting by human eyes to phases where the AI was fed external images again. In the end, 17 images out of more than 300 were selected from the work done by the AI. For more details, we invite you to read this tutorial.
It should be noted that Stable Diffusion is powered in particular by LAION, and the resulting graphic style is less realistic than Dall-E or less polished than MidJourney. However, it is a much more transparent AI where the code is accessible, unlike the products offered by OpenAI, and does not require the use of a platform such as Discord.
Image selection process
Once Amelia's corpus had been defined, it was then possible to place her in different contexts :
- Amelia became a major criminal : a cannabis or arms trafficker, and pornographic photographs of her were also leaked.
Where deepfake made it possible to integrate a face into an existing pornographic video, AI can compose a hyperfake.
To achieve a credible image, the process is still a long one. A multitude of improbable elements are hidden in the productions : guns become cameras, cannabis plants have strange proportions, bodies that are supposed to be naked are covered with rough patches.
Selection of three renderings by Amélia :
- Amelia soldier shooting with rifle machine gun, female army military fight training
- Amelia addicted using a cannabis bong, hold a join, addiction, realistic,
- No specific prompt, image generated from an image corpus
What could you make an individual do if you were accompanied by an AI system that had access to their personal information ?
- Amelia was able to get younger. Just like a filter that's currently trending on TikTok and lets users look younger in a believable way, Amelia goes back to being a kid. Then it's pretty easy to get her moving :
A series of screenshots from a deepfake of a young Amelia. The video still has many flaws, however.
The video produced is still full of glitches, but with enough time, the result could be more than credible. This is the current limit of the tool's performance, but it can be improved. Automations already exist with consumer applications that allow deepfakes to be created from video clips.
Could you use this technology to make yourself appear younger ? And vice versa, could a minor use AI to make themselves appear older ?
- Amelia now has an Instagram account. Just like Lil Miquela, she will soon be able to become a virtual influencer. The grid consists of photos created by Stable Diffusion and encoded by MidJourney, with descriptions written by ChatGPT. AI has become our personal content manager.
The image was generated by Stable Diffusion, while the description was written by ChatGPT. The AI was instructed to write the text for Amelia's first Instagram post based on her description.
Could our future virtual world be flooded with this kind of fake content ?
Amelia is now a fully developed persona : at first glance, the ease with which accounts can be created for her or she can be placed in a completely new situation would almost make you forget the initial difficulties in creating her.
However, this artificial Amelia is far from passing all identity checks, both by machine and by the human eye. Often, you feel a certain discomfort when observing AI productions, a similarity between them all and stereotyped elements, just like when you scroll through the retouched images and sanitized content on Instagram. Amelia, like all the characters proposed by ChatGPT, is happy, healthy, and in good health. When trying to create multiple profiles, it becomes increasingly necessary to guide the AI to prevent it from repeating the same stereotypes over and over again.
AI has fed on content from the internet that reflects society's biases, and it will be interesting to see how it develops in the future. Will the next step be standardization, where AI feeds on its own content ?
Summary
- Is there a (co-)pilot in the plane?
- Lock, Stock and Two Smoking Bairrels
This article is part of a series on generative AI, which begins here !
Illustration - LINC avec StableDiffusion
