The deep history of training datasets

In early 2021, OpenAI released the original DALL-E along with a paper describing the system. The paper revealed that DALL-E had been trained on roughly 250 million image-caption pairs scraped from the internet. The dataset itself was not released. OpenAI kept it private, citing commercial sensitivity and the difficulty of cleaning a dataset that large. Anyone who wanted to train their own image-text model would have to scrape the data themselves.

A German high school physics teacher named Christoph Schuhmann decided this was a problem worth solving. He gathered a group of volunteers on a Discord server for AI enthusiasts and proposed building an open replica of OpenAI's dataset. They called the project LAIONLAIONA German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion., Large-scale Artificial Intelligence Open Network. They were not researchers at any institution. They had no funding. They had no compute. What they had was determination and a willingness to do the unglamorous work of writing scripts that crawled 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. (the publicly available archive of the entire internet), extracted image URLs and their 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. captions, and stored the resulting pairs as a downloadable dataset.

Their first release, LAION-400M, came out in August 2021 with 400 million image-caption pairs. The total cost of producing it was about ten thousand dollars in server fees, covered by a combination of small donations and a contribution from Hugging FaceHugging FaceThe dominant repository for open-source AI models, including all the open-weight image and video models discussed in this document.. The next year they released LAION-5B, with five billion pairs, at the time, the largest freely available image-text dataset in existence. The funding for LAION-5B came partly from Stability AI, whose founder Emad Mostaque donated between nine and ten thousand dollars. That donation, more than any other single decision, made Stable Diffusion possible.

Stable Diffusion was trained on a subset of LAION-5B called LAION-Aesthetics, which contained about 600 million images that had been filtered down using an aesthetic classifier, a small neural networkNeural networkA mathematical function built from layers of simple operations, with millions or billions of adjustable numbers inside it, that learns a task by having those numbers tuned on examples. trained on human ratings of which images looked nice. Without LAION, there would have been no public dataset large enough to train Stable Diffusion. Without Stable Diffusion, there would have been no open-source generative imagery ecosystem. Without that ecosystem, there would be no FLUX, no ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org., no CivitaiCivitaiA community website for sharing Stable Diffusion and FLUX models, LoRAs, and embeddings., no Low-RankRank (LoRA rank)The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files. Adaptation (LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models.) scene, no Krea, no Flora. The entire bottom half of the field rests on Schuhmann's volunteer project.

Independent researchers have analyzed the contents of LAIONLAIONA German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion.-5B and the picture they paint is illuminating. About 47 percent of a 12-million-image sample came from just 100 domains. Pinterest alone accounted for 8.5 percent. The rest of the top sources were what you would expect from a scrape of the public web: WordPress, Blogspot, Flickr, DeviantArt, Wikimedia Commons, Shopify product pages, news sites, blog images. The 'global aesthetic' that Stable Diffusion learned in 2022 was, very literally, the aesthetic of Pinterest and of personal blog post header images, weighted by what an aesthetic classifier had decided was attractive. This is also why early Stable Diffusion outputs felt slightly stock-photo-ish, that is what most of the training data looked like.

This composition has significant downstream consequences. Models trained on web-scraped data inherit the biases of the web. They are more familiar with subjects that are overrepresented online (young attractive people, popular tourist destinations, common dog breeds, mainstream architectural styles) and less familiar with subjects that are underrepresented (elderly people doing ordinary activities, indigenous cultures, agricultural work, unfashionable cities). They also inherit the watermarks and stock-photo signatures of the sites they were scraped from, which is why early models sometimes generated images with semi-legible 'Getty Images' or 'Shutterstock' overlays in the corners. It took deliberate filtering work to remove these artifacts in subsequent model generations.

In December 2023, researchers at the Stanford Internet Observatory published a report that found more than a thousand verified URLs containing child sexual abuse material in LAIONLAIONA German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion.-5B. This was the original sin of the entire field. The dataset had been assembled by automated scraping with no manual review of every image, and even with extensive filtering, illegal content had made it through. LAION immediately took the dataset offline. In August 2024, after working with the Internet Watch Foundation, Human Rights Watch, the Canadian Center for Child Protection, and the Stanford team, LAION released Re-LAION-5B, a cleaned version of the dataset with the offending content removed.

The technical detail to understand: LAION never hosted images. It only hosted lists of URLs pointing to images on other sites. This is what gave them some legal insulation, and it is also what made the cleanup possible, they could update the URL list to remove problematic entries without taking down anything they were storing. But it also meant that any model trained on the original LAION-5B before the cleanup had absorbed those images during training, and there was no way to surgically remove them from the trained weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable.. The Stanford report did note that the presence of CSAM in the training set does not necessarily mean a trained model can produce CSAM-style outputs, modern models have additional safety filtering, but the fact that the data was there at all has shaped the field's relationship with web-scraped training data going forward.

The labs that came after Stable Diffusion have generally been more careful. Black Forest Labs has not disclosed exactly what FLUX is trained on, but the released models include extensive safety filtering. Adobe Firefly is the most extreme case in the other direction, it is trained exclusively on Adobe Stock content and other licensed material, which is why Adobe can offer commercial customers Intellectual Property (IP) indemnificationIndemnificationA vendor's promise to cover a customer's legal costs if the product triggers a lawsuit, offered by some licensed-data model providers as protection against copyright claims on generated output. that no other model provider can match. The trade-off is that Firefly's training data is much smaller and less diverse than the open scrape approach, which is why Firefly tends to feel less capable on niche or unusual prompts.

If you think image data is hard, video data is much harder. A video model needs not just images and captions but video clips with descriptions of what is happening in them across time. You cannot just scrape this from the web in the way you can for images. Most internet videos do not come with timestamped descriptions. The captions that exist (YouTube auto-captions, video titles, descriptions) describe what is being said or what the video is broadly about, not what is happening visually frame by frame.

The video research community has developed a few large datasets to address this. WebVid-10M, released in 2021, contained 10 million videos scraped from a stock video site with their accompanying titles as captions, a rough but usable starting point. WebVid was widely used until 2023 when it was retracted due to copyright concerns from the source site. HD-VILA-100M, released by Microsoft Research in 2022, contained 100 million video clips from YouTube with automatic speech recognition transcripts as their text. The transcripts were often unrelated to what was visible on screen, but at scale the noise averaged out enough to be useful.

The breakthrough that defines the current generation of video datasets is a technique called multi-teacher captioningMulti-teacher captioningA pipeline for building video datasets that runs several vision-language models on each clip to propose captions, then scores and keeps the best one, producing high-quality labels at scale., popularized by Snap Research's Panda-70M in February 2024. The idea is to take a large pool of raw videos, split them into short semantically coherent clips, and then run multiple different vision-language models on each clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. to generate candidate captions. The candidate captions are scored by a fine-tuned retrieval model that picks the best one for each clip. The result is 70 million video clips with high-quality automatically-generated captions, dramatically better than scraping.

Frontier video models like Sora, Veo, Kling, and Wan all use variations on this approach, with their own proprietary captioningCaptioningWriting the text descriptions that accompany training images. In LoRA training, whatever you caption becomes a controllable variable and whatever you omit becomes baked-in identity. pipelines. The exact numbers and sources are guarded secrets, but the general structure is: gather millions of videos from large libraries (YouTube for Google, TikTok-adjacent sources for ByteDance and Kuaishou, licensed stock for some Western labs), filter aggressively for quality and content compliance, run them through multiple VLM captioners, score and select the best captions, and use the resulting (clip, caption) pairs as training data. Companies that own large video platforms, Google with YouTube, ByteDance with TikTok, Kuaishou with their short-video platform, have a structural advantage that no startup can match. This is the single biggest reason the Chinese AI labs have caught up to and surpassed the Western labs in video generation.