A generative is a function. Something goes in, usually a sentence and sometimes a picture, and something comes out: an image, a paragraph, a few seconds of video. In between sits a very large set of numbers, called the , which were adjusted over weeks or months of computing until the output stopped being noise and started to look like the material the model was shown. The original Stable Diffusion, released in 2022, is a useful size to picture. Its image network holds about 860 million of these numbers and its text encoder another 123 million, and its makers say it runs on a graphics card with at least 10 gigabytes of memory. Those numbers are the model.
The weights got their values through , which is simple to describe and enormously expensive to do. The model is shown an example, makes a guess, is told how wrong the guess was, and then every weight is nudged a tiny amount in the direction that would have made it less wrong. Repeat that across hundreds of millions of examples. For an image model the examples are pictures with captions attached, and Stable Diffusion learned from subsets of LAION-5B, a public list of 5.85 billion image and caption pairs gathered from the web, of which 2.32 billion have English captions. Training happens once, at the maker's expense. Everything you do afterwards is , which means running the finished model to get an answer, and inference is the part billed to you every time.
The most useful thing to hold onto is that the model does not contain the pictures. It contains a compressed account of how pictures of that kind tend to go: what a shadow does under a bottle, what fabric does at a shoulder seam, which arrangements of pixels people tend to caption as golden hour. The arithmetic backs this up. There is no room in about a billion numbers to keep copies of hundreds of millions of photographs, so what survives training is the pattern and the individual picture is lost. Asked for something new, the model builds an image consistent with everything it learned, and that image did not exist until you asked.
Occasionally it does reproduce something it saw. In 2023 Nicholas Carlini and colleagues went looking for exactly that inside Stable Diffusion, and they searched on purpose among the 350,000 pictures that appeared most often in its training data, because a picture seen hundreds of times is the one most likely to be learned by heart. They generated 175 million images from those captions. They found 109 near-copies. So copying is rare and measurable, and it is most likely for widely reposted images, which is exactly the kind a brand or a photographer worries about. That is why the rights argument is real rather than theoretical, and why a court has already had to rule on what a model contains.
Two kinds of generative model do most of the work a designer will meet. Image and video models are mostly diffusion models, which start from random static and remove a little of it at a time, with your prompt pulling on every step, until a picture is left. Language models, the kind behind a chat assistant, write one small piece of text at a time, each time choosing a likely next piece given everything written so far. The two work in different ways. In both, the model has learned what is likely and produces its answer by drawing from that, step by step, and the newer models that read and produce text and pictures together are built from the same parts.
That drawing is the last idea you need, because it explains how these tools behave from one day to the next. Every output is a from a range of likely answers, and a number called the seed decides which one you get. Change the seed and the same prompt gives you a different picture; keep the seed, the model and the settings the same and you can usually get the same picture back. A working designer thinks of every input as a way of moving where in that range the model draws from. A prompt narrows it. A reference image narrows it much further, and a model trained on photographs of your own product narrows it most of all, which is why studios doing product work end up training one.
A real case
The High Court looks inside Stable Diffusion
Getty Images v Stability AI, High Court of England and Wales · 2025
What they did
Getty sued Stability AI over pictures scraped from its sites to train Stable Diffusion. By the trial Getty had dropped the claim about training, which had not happened in the UK, and the claim about outputs. What remained was the argument that the model itself was an 'infringing copy' of Getty's photographs.
What happened
On 4 November 2025 Mrs Justice Joanna Smith rejected it. Both sides' experts agreed that Stable Diffusion 'does not itself store the data on which it was trained', and she held that something which has never contained a copy cannot be an infringing copy. Getty won only a narrow trade mark point, over Getty and iStock watermarks that some early versions generated.
What to take from it
The court's account of a model is the one in this lesson: weights learned from pictures, with no pictures inside. The watermarks are the other half of it, since a model can still redraw what it saw over and over, logo and all. This is one first-instance ruling in England, and the training question itself was never decided.
The words, in plain language
- Model
- A very large set of numbers that encodes patterns found in training data, plus the code that runs them. It holds no library of pictures and cannot look anything up.
- Weights
- Those numbers. A model is its weights; everything else around it is software.
- Parameters
- How many weights there are. Often quoted as a headline figure and only loosely related to how good the thing is.
- Training
- Showing the model examples and nudging its weights after every wrong guess. Done once, by the maker, at great cost.
- Inference
- Running the finished model to get an answer. The part you pay for every time.
- Sample
- One draw from the range of answers a model finds likely. Two runs of the same prompt are two samples.
- Memorisation
- A model reproducing a training image almost exactly. Rare, and most likely for pictures it saw many times over.
Why it matters
This explains almost every frustration you will have. A model that has learned how pictures tend to go will give you a picture that is typical, and typical is the enemy of a brand. Ask it for a chief executive, an accountant or a nurse and it will give you the average of what its training data showed under those words, and those averages carry the prejudices of the material they came from. You cannot report this as a bug. It is the shape of the thing, and a person has to steer against it on every job.
Beginners make three mistakes with this, and each one costs hours. The first is treating the model like a search engine, asking it for a real product or a real person and trusting what comes back, when all it can produce is something that looks plausible. The second is rerolling. Two runs of the same prompt are two draws from the same range, so if the range is wrong for the job, fifty more runs give you fifty more wrong pictures; the fix is to change the input or change the model. The third is assuming the model learns from you as you work. Its weights do not change while you use it, so a correction you made in one session is gone in the next unless you repeat it, the tool saves it as a note, or someone trains the model on purpose.
It also explains what direction is for. If the machine produces the typical, the value a person adds is knowing what is not typical and insisting on it, in instructions specific enough to steer a statistical process. That is why a vague brief produces generic work with unusual reliability. Professionals also check where a model came from before it touches paid work. They read its model card, the page where the maker states what it was trained on, what licence it carries and what it is known to get wrong, and they look hard at any output that resembles a famous photograph or somebody else's logo.
The rest of this module opens up the machine one part at a time. You will see how a diffusion model turns static into a picture, what the compressed space it works in actually is, how words become numbers the model can steer by, and how attention lets every step of the picture consult your prompt. After that come the ways a model is trained further on your own product, the seed and the settings you will actually touch, and why these systems still miscount fingers and misspell labels. The last stretch covers how much a model can hold in mind at once, models that read pictures and text together, why published scores mislead, and what running all of this costs. None of it needs mathematics.
Elsewhere
Papers
- Type
- term
- Period
- ongoing