Notes

What happens to your image in untrainable.art

If you're going to run unreleased artwork through a tool that promises to protect it, it's fair to ask first where the image goes. With untrainable.art, it stays in your browser tab. Here's what happens to it there.

What your browser downloads. The page, and the first time you visit, a model (about 600 MB, from Hugging Face), so the tool can do its work on your machine. That's the trip in one direction. Your image never makes the trip the other way.

What stays in the tab. Everything about your image. When you drop it in, the browser reads it from your disk. You tell the tool what the image shows and what you'd like a model to mistake it for (the presets include painting-to-photograph and illustration-to-sculpture). It then adds noise based on the Nightshade research, step by step, in the tab. The default strength is 10 out of 255 and the default number of iterations is 100, so the change to any one pixel is small. When it's done you get the original, the poisoned copy, and the noise on its own at ten times strength so you can see what was added. The download is generated on your machine too.

The settings you might change. Turning the strength up makes the noise easier to see. More iterations take longer. Under Advanced there are options to concentrate the noise in mid-range frequencies, where it's harder to notice, and to embed a "not for AI training" flag in the file's metadata.

Why your browser matters. On Chrome or Edge with WebGPU, the tool runs at full quality. In other browsers it falls back to the CPU and runs fewer iterations, so it finishes faster but the noise is weaker. If you're protecting something that matters to you, use a WebGPU browser.

How to check for yourself. Open your browser's developer tools, go to the Network tab, then reload the page and process an image. You'll see the page come from untrainable.art and, the first time you visit, the model and the code that runs it come from Hugging Face and jsDelivr. You'll also see Plausible, our analytics, load its script and send two small requests: one when the page opens, and one when an image finishes processing. None of them carries your image.

What it can't promise. untrainable.art is built on Nightshade, research from the Glaze Project at the University of Chicago. In the tool's own words, models that train on poisoned images "learn corrupted associations." We don't claim more than that. Nobody can guarantee what a particular model will do with a particular image.

untrainable.art is free to use. If you make things for creators, we'd like to hear from you.