What the models were trained on

An image model learns by looking at pictures paired with words. During training it is shown hundreds of millions or billions of images, each with a caption, and it slowly adjusts itself until it can connect a phrase like 'linen shirt in morning light' with the patterns of pixels that tend to go with it. It keeps no folder of pictures. It keeps what it learned from them, which is why the question of what it was shown matters so much. A model that saw mostly English captions understands English prompts best, and a model that saw millions of product shots makes convincing product shots. What it can do, and what it gets wrong, both trace back to the training data.

For most large image models, that data came from the open web. The best documented example is LAION-5B, a public dataset released in 2022 that lists 5.85 billion links to images alongside their captions, gathered automatically from web pages across the internet and filtered with another model, called CLIP, that scores how well a caption matches its picture. It holds links and text instead of the images themselves. Over two billion of its pairs are in English. The first versions of Stable Diffusion were trained on subsets of that English portion, and their model card says so. So a great deal of work by living illustrators, photographers and designers went into training. Nobody asked or paid them. The papers themselves say where the images came from.

A scraped dataset holds whatever was online, including the worst of it. In December 2023 a researcher at the Stanford Internet Observatory found around a thousand links in LAION-5B pointing to suspected child sexual abuse material. LAION took the dataset offline on 19 December 2023, and in August 2024 it released a cleaned version, Re-LAION-5B, with 2,236 suspicious links removed. Stable Diffusion's card adds its own warnings. It says the data contains adult material and was not de-duplicated, so the model shows some memorisation of repeated images, and that 'white and western cultures are often set as the default'. In 2023 researchers led by Nicholas Carlini extracted over a thousand training examples from models including Stable Diffusion, from photographs of real people to company logos.

A minority of makers built their models the other way and say so. Adobe states that its own Firefly models were trained on licensed content such as Adobe Stock, along with openly licensed work and public domain material. Getty Images' generator, built on Bria's Fibo Lite, was trained on images owned or licensed by Getty and named partner libraries, and its model card says none of it was scraped from the internet or made by another generator. These models trade some range for a clean answer. Getty's training data stops at October 2023. Several providers now offer legal indemnity to enterprise customers, which is a commercial admission that the question is real.

Between those two groups sits a large middle of models that publish their architecture and their licence and say nothing about the images. The model cards for SDXL base 1.0 and FLUX.1 [dev] are examples. That silence is itself an answer. The argument about all of this is still open, and both sides deserve a fair hearing. Supporters of training on public web data say a model learns patterns the way a student learns from looking, and that copyright covers particular works while leaving general lessons drawn from them free to use. Artists and photographers reply that their work was copied in full, without consent or payment, to build a product that now competes with them for the same commissions. Courts in several countries are hearing versions of both arguments.

Opting out is the main tool artists have, and it is weak. An opt-out only affects training that happens after it, and it works only where the trainer checks for it. In the European Union, the law lets rights holders reserve their work from commercial text and data mining, but a German court ruled in December 2025 that a reservation written in ordinary words on a website did not count, because software could not read it. So a working professional who publishes a portfolio puts the reservation where machines look, such as a robots.txt file or the TDM Reservation Protocol. They keep a dated record of when it went up, and they treat it as a request that some trainers will ignore.