What the training run actually consisted of

Billions of pairs of picture and caption

The open image models of the past few years were trained on web-scale datasets built the same way: crawl the web, keep every image that has alt text, filter, and store the pairs. The best-documented is LAION-5B, released in 2022, with roughly 5.85 billion image-URL and caption pairs derived from Common Crawl. LAION itself stores no images — only links and text — which is part of how it was assembled at that scale and also why parts of it decay as pages disappear.

Two things about this deserve to be said plainly.

First, alt text is not description. It is what a webmaster typed to satisfy an accessibility checker or a search engine. A large share of it is a file name, a product code, a stock-photo watermark caption, or a keyword string. The model's idea of the English language is this corpus, not literature.

Second, the assembly was not curated in any meaningful sense. In December 2023, researchers at the Stanford Internet Observatory found that LAION-5B's index contained links to child sexual abuse material. The dataset was withdrawn, cleaned against known hash lists, and re-released as Re-LAION-5B in 2024. This is not a footnote. It is the clearest available evidence about what "scraped from the open web" means as a sourcing practice, and it is why later datasets from serious labs are filtered far harder and documented far less.

Why this dictates how you prompt

The strange conventions of image prompting are downstream of the caption corpus.

"Trending on ArtStation" worked because those words appeared in the alt text and surrounding metadata of a particular kind of polished digital illustration. "35mm, f/1.8, bokeh" works because photographers write their settings into captions, and those settings correlate with a shallow-focus look. "Unreal Engine" works for the same reason. None of these are instructions the model understands. They are strings that co-occurred with a visual style.

That also explains the failures. The model has weak coverage of anything the web does not caption well: specific tools, regional dress with names that vary by language, technical diagrams, anything photographed mainly by people who do not write English alt text. Ask for a lungi or a sarpech or a specific loom and you will get something adjacent, because the caption density is thin. Ask for a wedding dress and you will get white satin, because the caption density is enormous and Western.

When a prompt fails, the first question is not "how do I phrase this better" but "would anyone have written this caption on the web". If the answer is no, phrasing will not save you and you need a reference image or a fine-tune.

Newer training, and what changed

The most visible recent improvement in prompt following came from recaptioning. Rather than train on the original alt text, labs pass every training image through a vision-language model that writes a long, literal description, and train on that. It is expensive and it works: models trained this way follow sentences with spatial relations, counts and multiple attributes far more reliably, because their captions during training actually contained those things.

It has a side effect worth knowing. Models trained on machine-written captions respond best to prompts that look like machine-written captions — full sentences, literal, describing what is in the frame. The old tag-soup style degrades on them. If a new model seems worse than an old one on your existing prompts, this is usually why, and the fix is to rewrite the prompt as a description rather than a keyword list.

What this means for the ownership argument

Hold this lesson next to the legal module later in the course. Every serious dispute about generative images starts here: billions of copyrighted works were downloaded and used as training input, in most cases without a licence and without a workable way to opt out, on the argument that training is a transformative analysis rather than a reproduction. Courts in different countries have begun to answer that differently. Whatever you conclude, the factual base is not in dispute and is worth knowing precisely: web scrape, alt text, billions of pairs, no consent mechanism at collection time.