RN
Start with the thing that surprises people first: ask one of these systems for the same image twice and you get two different pictures. Not variations on a stored photograph — two separate images, neither of which existed before you asked.
Nothing is being retrieved. That single fact explains most of what is strange about the category, including the good parts and the parts that should worry you.
What the machine is doing
The dominant approach is called diffusion, and the training procedure runs backwards from what you would expect.
You take an enormous set of images and systematically destroy them — adding random noise in steps until nothing is left but static. Then you train a network to undo one step of that destruction: given a noisy image, predict what the slightly-less-noisy version looked like. Do that across enough images and enough noise levels and the network learns, in effect, what it takes for a picture to become more like a picture.
To generate, you hand it pure noise and run the learned denoising repeatedly. Structure emerges out of the static because the model is pulling every step toward the kind of image it was trained on. A prompt steers that pull: the text is encoded into a vector, and the denoising is conditioned on it, so each step nudges the image toward the region of possibility the words describe.
An earlier generation of tools used GANs, where a generator and a critic network train against each other until the generator's output fools the critic. They produced striking results in narrow domains and were notoriously unstable to train. Diffusion largely displaced them for open-ended generation because it scales better and handles text conditioning more gracefully.
Why prompts behave the way they do
A prompt is not an instruction and the model is not parsing it as a sentence. It is a bias applied to a search through possibility space, which has three practical consequences.
Some words carry far more weight than others. Terms that were common and consistent in the training captions steer strongly. Terms that were rare or inconsistently used barely register. This is why one adjective can transform an image and another does nothing at all, and why the effective vocabulary of a generator has to be learned by experiment rather than read off a list.
Composition is weaker than content. These models are much better at what is in an image than at where things are relative to each other. Spatial relationships, counts, and any instruction of the form "A on the left, B on the right" fail routinely, because the training signal for them was thin.
Negation is not reliable. Telling a model to exclude something often makes it more likely to appear, since the concept is present in the conditioning either way. Most interfaces solve this with a separate negative-prompt field, which works by steering actively away rather than by understanding the word "no".
The seed is the other half. It is the specific random noise the process starts from. Same seed plus same prompt plus same settings gives the same image; change the seed alone and you get a different picture matching the same description. Every interface that shows you a grid of four options is showing you four seeds.
Where it still breaks
The artefacts have moved but not disappeared, and they cluster in predictable places.
- Hands and fingers. Improved substantially, still the most common tell. Count them.
- Anything with a long-range dependency. A necklace chain that changes thickness behind the neck, a strap that goes into a shoulder and does not come out, a tattoo that reads differently on each arm. The model works locally and has no persistent object model.
- Written text. Better than it was, still unreliable at small sizes and in reflections.
- Backgrounds under low attention. Objects behind the subject merge, and architecture bends. Look at door frames.
- Physical implausibility that reads as fine at a glance. Limbs at angles that no joint permits, weight distributed in ways no body would hold. This is the class of error that a still image hides and motion exposes.
Video is a categorically harder problem, because the model must be consistent not just within a frame but across frames. Faces drift, clothing patterns reshuffle, and things in the background appear and disappear. The current generation manages short clips at increasing fidelity; sustained, continuous, coherent scenes remain difficult in ways that are obvious the moment you watch for them.
The training-data question
Every one of these models learned from photographs of people, and almost no commercial generator publishes what those photographs were.
That matters for two separate reasons and they get conflated constantly.
The first is a rights question: whether images were licensed, scraped, or taken from adult productions whose performers agreed to a shoot and not to a dataset. The second is an output question: a model trained on a body of work carries its statistical regularities forward, which is why generators have such strong and consistent default aesthetics — the sameness people notice across an entire platform is the training set showing through.
There is also a specific capability worth understanding plainly. Lightweight fine-tuning methods let a general model be adapted to a particular subject from a modest number of images, cheaply, without retraining anything from scratch. That is a legitimate tool for consistent original characters. It is also precisely the mechanism behind synthetic imagery of real people who never agreed to any of it, and the technique does not distinguish between the two uses. Nothing about the maths cares whose face it was.
So what is it good for
Honestly assessed, the strength is not realism. It is specificity — combinations that no one filmed because no market for them existed, aesthetics that are not in anyone's catalogue, and iteration where you adjust and regenerate rather than search and settle. That is a real difference from a library of recorded material, and it is the reason the generator category is a distinct thing rather than a novelty attached to the existing one.
The weakness is that everything is average by construction. A model produces the most probable image consistent with your words, which means the interesting failure — the odd angle, the unrepeatable moment, the thing a person did that nobody planned — is exactly what it cannot give you. The output is fluent and it is never surprising.
Whether that trade appeals is a matter of taste. What it is not is a replacement, and the AI category reads better as a new kind of thing sitting alongside the old kind than as the future arriving to displace it.