Glaze is not really in this comparison, and that is the most useful thing to know
Glaze and Fawkes both come from the University of Chicago SAND Lab, both add a perturbation that is meant to stay below the threshold of ordinary human perception, and both get filed under the same heading in articles about protecting yourself from AI. That is where the similarity stops.
Glaze protects an artist’s visual style. The perturbation is designed to interfere with an image-generation model learning and reproducing how someone paints, so that training on their portfolio produces something that does not look like their work. It is available as a desktop application and as WebGlaze, a free invite-only hosted version processed on the project’s own GPU servers.
It does nothing for face matching. If your problem is that a public headshot is turning up in a face-search engine, Glaze is the wrong tool, and running your photo through it will not help no matter how many times you do it.
Fawkes and LowKey are the comparison that actually exists
Both are academic projects and both attack the same step. A face-recognition system converts a face into an embedding — a list of numbers describing its geometry — and decides whether two photos are the same person by comparing embeddings. Fawkes and LowKey both compute a perturbation that makes the model extract a substantially different embedding from your photo than it otherwise would, while leaving the image looking like your photo to a person.
The commonly reported difference between them is visibility. LowKey’s alterations tend to be more noticeable than Fawkes’. Whether that is a problem depends entirely on what the photo is for: on a professional headshot it matters a great deal, and on a photo where you care more about protection than polish it matters much less.
There is no single best tool here, and anyone presenting one as the winner is skipping the actual axis. Protection strength and visual fidelity trade against each other, both tools expose that tradeoff through settings, and where you want to sit on it is a judgement about the specific image rather than a fact about the software.
The limitations are shared, and they matter more than the choice
None of them is retroactive. Cloaking changes the file you cloak, at the moment you cloak it. Copies of your face already indexed by a face-search engine are completely unaffected, and getting those out of results is a separate process of opt-out and removal requests to each engine.
Recompression degrades all of them. Platforms resize, re-encode and strip metadata from photographs on upload, and the perturbation lives in exactly the fine detail that compression discards. This is a known limitation of the whole category rather than a weakness of any one tool, and it is the one that gets mentioned least.
One clean copy undoes the work. If the unprotected original is published anywhere — an older profile, a press page, a colleague’s post — an indexer has an uncloaked reference and the cloaked version does not help.
And none of it is permanent. Detection is plausible in principle, because the perturbations follow patterns a system expecting them can look for, and reversal is an active area of research. The honest framing, which is the one the researchers who built these tools use themselves, is that cloaking raises the cost and unreliability of matching rather than guaranteeing anything.
Choosing, in practice
If you are protecting a professional headshot and the image has to look untouched, the lower protection modes of Fawkes are the conventional answer, and they are what this service runs.
If you are comfortable in a terminal and have more than a couple of photographs to handle, run Fawkes yourself. It is free, it is from the researchers who invented the technique, and your photo never leaves your computer — which on the privacy axis specifically is strictly better than any hosted service, including this one. Paying for a browser upload makes sense when you want one photo handled in about a minute without installing a Python environment and downloading model weights.
If you are an artist protecting a style rather than a face, the tool is Glaze, and it is not ours.