AI image generation: how does it work?
Today, artificial intelligence tools are capable of creating original images from simple textual descriptions, opening up unprecedented opportunities for marketing, communication, and media.
These images are generated using two technologies: GANs (Generative Adversarial Networks) and diffusion models.
- GANs (Generative Adversarial Networks)
GANs were invented in 2014 by Ian Goodfellow, who currently works at DeepMind, Google's subsidiary specializing in artificial intelligence. Their principle is based on a pair of algorithms that work in opposition:
- The Generator creates images from random data.
- The discriminator evaluates whether the image produced is "real" (i.e., similar to images from real data) or "fake."
Imagine a forger (the generator) trying to create convincing works of art, and an art expert (the discriminator) tasked with detecting the fakes.
At each iteration:
The forger improves his techniques to deceive the expert, and the expert refines his critical eye to unmask the fakes.
These iterations push the generator to produce increasingly realistic images. GANs are behind many innovations, particularly in the creation of hyperrealistic portraits and the generation of fictional faces used in digital marketing.
2. Dissemination Models:
Diffusion models, which are now the most widespread and have almost completely replaced GANs, have recently been popularized by tools such as DALL·E 2, Stable Diffusion, Midjourney, and many others. They operate on a completely different principle, which consists of creating order out of chaos.
- The model takes an image and adds random noise (like "grain" on an old photo) until the image is completely unrecognizable.
- The model then learns to do the reverse: gradually remove the noise to "reconstruct" the original image.
- Once trained, the model can start with a completely noisy image and generate a new image from a simple textual description.
It's a bit like having a sheet of paper covered in ink blots, and the AI knows how to reveal a specific drawing by gradually erasing the blots.
Today, diffusion models dominate the generative image creation scene thanks to their flexibility and ability to produce complex works from simple text prompts.
What uses are there for marketing, communication, and media?
These technologies offer multiple opportunities.
- rapid creation of advertising visuals, custom illustrations, and creative concepts.
- Generation of illustrative images for articles, attractive YouTube thumbnails, or even automated videos.
- Customization of graphic content for social media without systematically relying on traditional image banks.
Ethical issues
However, these tools do not replace human creativity; they simply enable users to work faster and respond with greater precision, provided that they have specified their requirements clearly.
While AI enables productivity gains and increased creativity, it also raises ethical questions about its potential use in spreading misinformation (deepfakes) and calls into question the entire body of case law on intellectual property rights for AI-generated works.
Last August, a group of organizations representing creators and copyright holders in Europe published an open letter to the European Commission "calling for effective and meaningful implementation" of the Artificial Intelligence Act (AI Act), which came into force on August 1.
For their part, faced with legal uncertainties, many advertisers are reluctant to use these models, even if they authorize their agencies to use them under certain conditions.