Image generator

An image generator is a program to which you describe, in plain language, what you want to see. For example, you type: “a red fox in the snow, photographic”. A few seconds later, an image appears that shows exactly that. This image did not exist anywhere before. It was not searched for on the internet, but newly calculated by the program. This is possible because the program has previously seen millions of images along with their captions and has learned connections between words and image content from them.

What changes for photographers and graphic designers

Creating images was manual work for centuries. You needed a camera, a brush, or many hours at the computer. An image generator pushes these costs down to almost zero. An advertising agency that used to organize a photo shoot with a team and studio now first tries out fifty generated drafts. This is changing entire professional fields, especially for simple commissioned images such as stock photos or product visualizations.

At the same time, a tangible legal problem arises. The training images largely come from the open internet, often without the consent of the creators. Photo agencies and artists have therefore filed lawsuits in several countries. Whether training with other people’s images is permitted has not yet been conclusively clarified. For companies, this is a real risk, which is why some providers now only train with licensed image collections.

The third point concerns everyone: photos are no longer proof. A convincing image of an event that never took place costs nothing today. Experts call such fabricated images deepfakes. Newsrooms and courts must adjust to the fact that an image alone proves nothing.

From noise to finished motif

Most image generators today work with a method called diffusion. The basic idea is best explained backwards. During training, a real photo is taken and image noise, i.e. random color speckles, is gradually poured over it. After enough steps, nothing remains of the photo. The program now learns to reverse each of these steps: from a slightly noisy image, it is supposed to predict the somewhat cleaner one.

During generation, the program then starts with pure random noise. Step by step, it removes noise until a clear image remains. This is roughly like someone carving a block of marble that initially has no shape at all. The typed-in text determines which figure emerges. This text is first translated into numbers and guides the decision at every step as to which noise is removed.

That is why the same text never produces exactly the same image twice: the starting state is random. And that is why image generators have long struggled with details that require a fixed rule. Hands with five fingers or legible writing were for years the typical weak point, because the method reproduces patterns rather than counting.

From Midjourney to the smartphone camera

The best-known providers are called Midjourney, DALL·E from OpenAI, Stable Diffusion, and Google's Imagen. Stable Diffusion is particularly interesting because its program code is openly available. Anyone who owns a powerful graphics chip can run the model on their own computer, without a paywall and without content filters.

In everyday life, you often encounter the technology without it being called by that name. The “magic eraser” function in smartphone photo apps removes distracting passersby and fills the gap with generated pixels. Video conferencing programs generate backgrounds, online shops display clothing on generated models. In news reports, the term usually comes up in two contexts: copyright lawsuits and fabricated images surrounding elections or crises.

A common misconception, by the way, is that an image generator assembles existing photos like a collage. That is not true. What is stored are not images, but only numerical values that describe statistical patterns. A model with several billion of these values takes up only a few gigabytes, even though it learned from billions of images. Only for very frequently occurring motifs can it reproduce templates almost verbatim.