Where the training data actually comes from

The starting point for almost every dataset is Common CrawlCommon CrawlA nonprofit that has continuously crawled the public web since 2008 and publishes the archive for free, the raw source most image training datasets are extracted from., a non-profit that has been continuously crawling the public web since 2008 and publishing the results as a free downloadable archive. Common Crawl produces a fresh snapshot of about 250 terabytes of web pages every month. It does not focus on images specifically, it just downloads web pages, but every web page contains image links, and you can write a separate program that reads the Common Crawl archives, extracts every image URL, downloads the image, and saves it alongside whatever text was associated with it on the page.

The text association is the trick that makes this work for training. HTML has an attribute called 'alt', short for alternative text, that web developers are supposed to fill in with a description of each image for accessibility purposes (so screen readers for blind users can describe the image aloud). The alt-textAlt-textThe description attribute web developers add to images for accessibility, which became the enormous free caption corpus that made text-to-image training possible. was never meant to be training data, but it turns out to be an enormous, free, unstructured caption corpus. A photo of a golden retriever on a real estate listing might have alt='golden retriever sitting on grass in front yard.' A diagram on a science blog might have alt='figure 3, the citric acid cycle.' Multiply this across the public web and you have hundreds of millions of image-caption pairs that nobody had to manually label.

The most famous dataset built this way is LAIONLAIONA German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion.-5B, a nearly six-billion-pair scrape that became the training foundation for Stable Diffusion and most of the open-source image models that followed. The full human story of how LAION got built, who built it, how much it cost, and how a Stanford team discovered something terrible in it in late 2023, is told in the training-datasets chapter. For the purposes of this chapter, the thing to understand is that LAION and its descendants are all variations on the same pipeline: crawl the web, grab every image URL plus whatever text is near it, run automated quality and safety filters, release the filtered list. This is how the field got its training data, and it is also why the legal status is so contested.

The legal status of training models on scraped web data is the most contested question in AI right now and it is genuinely unresolved. The argument for legality is that the model is not storing or redistributing copyrighted images, it is learning statistical patterns from them, which is comparable to a human artist studying other artists' work, and that this falls under fair useFair useThe US legal doctrine that permits some unlicensed use of copyrighted work, and the central defense labs raise for training models on scraped data. Its application to AI training is still unresolved. in the United States and analogous doctrines elsewhere. The argument against legality is that the model can reproduce training images on demand for some inputs (a phenomenon called memorizationMemorizationWhen a model can reproduce specific training examples on demand rather than only learning general patterns, a central concern in the copyright debate.), that even when it does not directly reproduce them it commercially exploits the original artists' creative labor without consent or compensation, and that scraping the images in the first place may violate the terms of service of the websites they were taken from.

There are now several major lawsuits working through the US and European courts. Getty Images is suing Stability AI in both the UK and the US. A class of visual artists is suing Stability, Midjourney, and DeviantArt. The New York Times is suing OpenAI (though that case is about text, the legal principles overlap). Disney and Universal are suing Midjourney over character outputs. None of these cases have produced final rulings as of early 2026, but the consensus expectation is that the eventual outcome will involve some combination of (a) clarifying that training is legal under fair use but with conditions, (b) requiring opt-out mechanisms for rights holders, and (c) creating new licensing frameworks for commercial models. The exact shape of the resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. will significantly affect which models can be trained on what data going forward.

In response to the legal uncertainty, some labs have built models trained entirely on licensed dataLicensed dataTraining data used with the permission of and usually payment to its rights holders, the basis for models marketed as commercially clean., where every image was acquired through a paid agreement with the rights holder. Adobe Firefly is the most prominent example, trained on Adobe Stock plus public domain content. Getty Images released its own generative model trained only on its licensed photo library. Shutterstock has a similar product. The pitch to commercial customers is that these models carry no legal exposure, you can use the outputs in advertising, branding, and commercial work without worrying about an artist suing you in three years.

The trade-off is that licensed-data models are noticeably worse than open-data models on most quality benchmarks, because the open data is much larger and more diverse. Adobe Firefly's outputs are competent but conservative; the model has not seen as many edge cases as FLUX or Imagen and it shows. The gap is closing, but it has not closed. The strategic bet that Adobe and others are making is that legal clarity matters more than raw quality for enterprise customers, and that the next round of model improvements will narrow the quality gap to a point where the legal clarity tips the decision. So far, that bet looks reasonable but unproven.

The newest direction in data collection is to skip the web entirely and generate synthetic training data using existing models. The technique works in two ways. The first is to use a frontier image model to generate millions of new images with controlled prompts, then use those images to train a smaller or specialized model. The second is to use a large language model to generate higher-quality captions for existing scraped images, since the alt-textAlt-textThe description attribute web developers add to images for accessibility, which became the enormous free caption corpus that made text-to-image training possible. from the web is often terse and generic, while a language model can produce a long, detailed, accurate caption. Both techniques are now standard practice at the major labs.

The interesting recent finding is that synthetic dataSynthetic dataTraining data generated by existing AI models, such as model-made images or machine-written captions, increasingly used to supplement or replace scraped web data. can dramatically improve model quality even when the synthetic data is generated by the same model you are training. Black Forest Labs has been open about using synthetic captions extensively in FLUX training. OpenAI has acknowledged synthetic data plays a role in DALL-E and Sora. By 2027, the consensus expectation is that the largest models will be trained on a hybrid of scraped web data, licensed dataLicensed dataTraining data used with the permission of and usually payment to its rights holders, the basis for models marketed as commercially clean., and synthetic data, with the synthetic component growing each year. The implication is that the long tail of weird, low-quality, messy web images that defined the LAIONLAIONA German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion. era is slowly being replaced by curated, high-quality, model-generated training data, and that shift will reshape the field again.