A look at 12M+ of the 2.3B images used to train Stable Diffusion reveals ~47% came from only 100 domains; the largest number, or 8.5%, originated from Pinterest
laion-aesthetic-datasette Use Datasette to explore … Tweets: Andy Baio / @waxpancake : What images are in the massive dataset that trained the Stable Diffusion text-to-image AI model? @simonw and I made a tool for you to explore/search a subset of it: 12 million image/caption pairs out of over 2 billion that it was trained on. https://waxy.org/... Simon Willison / @simonw : @BlancheMinerva It's size. It's not just the cost of hosting something that big, it's also the time needed to prepare it. I'm just about patient enough to crunch the data for 12m records, but I'm not ready to spend many hours (if not days) running and debugging scripts against 100x the data. Ian Linkletter / @linkletter : Very sharp work here. I'm not surprised that Pinterest is the biggest source (in the sample) for image data. https://twitter.com/... Jackson Palmer / @ummjackson : Interesting look into which sources were scraped for the Stable Diffusion AI model. https://waxy.org/... Stella Rose Biderman / @blancheminerva : @simonw Whats the reason for using a sample instead of the whole thing? Compute limitations? Too much results for a human to sift through? Something else? If you just need more compute I would be happy to help by providing however much you need. Simon Willison / @simonw : Andy and I worked together on a tool for searching and browsing a sample of the images that were used to train Stable Diffusion - it's a useful way of getting a better understanding for how these things actually work https://twitter.com/... @vectorpoem : To be super clear, the problems most of us are identifying with this tech aren't that “all the images on the internet are... on the internet!”, it's that people created and unleashed a tool that does a thing w/ others' work at a scale it wasn't possible to do before, w/o consent. https://twitter.com/... @greyopals : Artists, it's worth looking into this! I did indeed find some pieces of mine, which was disheartening to say the least—I had assumed that since I'm relatively new to this industry, nothing of mine would be on there....?! but apparently, there is no way to hide🙃 https://twitter.com/... https://twitter.com/... @xemu : Lord give me the confidence of tech companies just deciding they can use 2.3 Billion images they have no commercial rights to in order to launder them through a ML system. https://twitter.com/... Andy Baio / @waxpancake : Several people have asked if artists consented to be in the training dataset or if there's a way to remove your work from it. The answer to both is no, an issue that I wrote about in my previous post. https://waxy.org/... @michael_nielsen : Thoughtful discussion of the ethics of using AI models trained on artists' artworks. Something I wonder about: artists, of course, learn from & are influenced by each other. But I don't refuse to view art because of this. Not sure where I should draw the line. https://twitter.com/... Simon Willison / @simonw : Andy and I have been collaborating on a deep dive into the openly available training data that was used for Stable Diffusion - here's Andy's detailed write-up on what we've figured out so far https://waxy.org/... https://twitter.com/... Karla Ortiz / @kortizart : 1/ Very interesting tool that checks Stability Diffusion's dataset. I searched my name and saw plenty of my work. Note from the authors " this is only a small subset of the total training data: only 0.5% of the 2.3 billion images that it was first trained on. https://twitter.com/... Timnit Gebru / @timnitgebru : Would be great if there was an effort to create datasheets, data statements or related documentation for the entire dataset. https://twitter.com/... Andy Baio / @waxpancake : @simonw The biggest single source for the Stable Diffusion image data in our sample was Pinterest, with 8.5% of the images scraped from their CDN, over a million URLs. Karla Ortiz / @kortizart : Here's the tool to check out if your work is in a tiny subset of one the three massive databases for Stability Diffusion: https://laion-aesthetic.datasette.io/ ... The article explaining the tool is extremely interesting as well: https://waxy.org/... Great work by @waxpancake and @simonw !! Neil Blevins / @artofsoulburn : Ever wonder if you're work has been included in an ai dataset? This site lets you search at least a portion of the dataset for stable diffusion. I'm surprised how primitive the descriptive text is that's paired with the images. Other models I believe use more precise data. https://twitter.com/... Daniel Danger / @tinymediaempire : 17 of my images were used in the training dataset of one of the biggest AI “art” generators. did i say this was ok? i did not. i absolutely did not. https://twitter.com/... Andy Baio / @waxpancake : We searched the image dataset for over 1,800 popular artist names, which you can browse and search for yourself. Of the top 25 most-mentioned artists in our sample, only three are still living. The most frequent? THOMAS KINKADE, PAINTER OF LIGHT™ https://twitter.com/...
Context & Ripple Effects
Andy Baio and Simon Willison have published laion-aesthetic-datasette, a searchable Datasette tool exposing a 12M-pair sample of the 2.3B images behind Stable Diffusion — the first practical way for outsiders to inspect what the model actually learned from. The headline finding is concentration: ~47% of sampled images trace to just 100 domains, with Pinterest alone at 8.5%, meaning a handful of sites function as de facto suppliers to the model.
The tool lands in the middle of an escalating consent fight: [[a:984092|artists had already discovered their work was in the training set without notice, payment, or any way to remove it]], and the same LAION corpus later drew scrutiny when researchers found thousands of CSAM instances in it (Stanford's count) before the full 5B+ dataset was pulled from download (LAION-5B's unavailability). A browsable sample turns those abstractions into something individual creators can query by name.
First-order effects
- Artists and rights holders can now search the sample to check whether their own work is inside Stable Diffusion's training data — turning 'I suspect my work was scraped' into a verifiable lookup, with no removal mechanism attached.
- Pinterest and the other ~100 domains supplying nearly half the sampled images are revealed as load-bearing sources of the model, putting their hosting practices under a scrutiny they did not opt into.
Second-order effects
- The confirmed no-consent, no-removal structure of the dataset hardens the artists' position in the dispute with Stability AI, since the tool documents exactly whose work is implicated rather than leaving it anecdotal.
- High-supply domains like Pinterest face pressure to decide whether they permit this kind of bulk reuse — a question that extends to every platform whose CDN-hosted images quietly feed scraped corpora.
Third-order effects
- If scraping-built datasets keep getting opened to inspection after the fact, consent, provenance, and takedown tooling stop being optional extras and become baseline requirements for anyone assembling web-scale training corpora.
- Concentration findings like this one suggest future licensing fights will be negotiated with a small set of dominant source domains rather than millions of individual creators — reshaping who has leverage in AI training-data deals.
The trend: Web-scale AI training sets assembled by silent scraping are being forced open to public inspection, converting dataset composition from a technical detail into a consent and accountability battleground.