"AI tool" almost always means the same thing under the hood: your file gets uploaded to someone else's server, a model runs there, and a result comes back. That's true for most background removers, most object detectors, most anything with "AI" in the name — the intelligence lives in the cloud, and your data has to travel there to use it. Tuliq's 🔵 Local AI tools work differently: the model itself is downloaded to your browser, and every photo you process after that runs entirely on your own device. Nothing is ever uploaded, not even to Tuliq.
What "runs in your browser" actually means here
This isn't a simplified or "lite" version of a cloud AI feature — it's a real trained model, the same kind used anywhere else, just packaged to run through WebAssembly (a way to run compiled, near-native-speed code directly inside a browser tab) instead of on a server. The sequence is:
- The first time you open a local-AI tool, your browser downloads the model file — a few hundred kilobytes to a few megabytes, depending on the tool.
- Your browser also loads a small WebAssembly runtime capable of executing that model.
- From then on, every image you process runs through that model on your own device, using your own CPU (or GPU, where the browser supports it). No network request happens during processing — the photo never leaves your machine.
That one-time download is the honest tradeoff: the very first use of a local-AI tool takes a moment longer than a purely instant 🟢 browser tool, while the model and runtime load. Every use after that — on that page, in that browser — is immediate, since both are cached.
Why this is a real distinction, not a technicality
A cloud AI tool's privacy promise is only ever "we say we don't keep it" — the file still has to leave your device to get processed at all, so you're trusting a policy, not an architecture. A local-AI tool doesn't have that step to trust in the first place: there's no server in the loop for the actual processing, so there's nothing to send and nothing for a policy to promise about. This matters most exactly where people are least willing to upload something to begin with — a photo with a person's face in it, an image tied to something private, a file from a context where "some company's server processed this" is a real concern, not a paranoid one.
The honest tradeoff runs the other way too: a model that has to fit in a browser download and run on ordinary consumer hardware is necessarily smaller than what a cloud service can throw a data-center GPU at. Tuliq's local-AI tools are chosen for being genuinely good at one clear task, not for matching the absolute state-of-the-art a massive server-side model could reach.
Where this shows up on Tuliq today
Object Detector uses a general-purpose detection model (~7MB) trained on about 90 everyday object categories — people, vehicles, animals, common household items — and returns a bounding box, a class name, and a confidence score for each one it finds. Its output plugs directly into Tuliq's existing annotation pipeline: see how to use AI object detection to speed up annotation.
It was chosen deliberately for having a permissive open license suitable for a public site, not simply for being the most accurate option available — a detail covered in the object detection workflow guide, since that's where the choice actually matters most. It's also a real example of the accuracy tradeoff mentioned above cutting the other way: a browser-sized model is good at what it's built for, not at everything — a general in-browser tool is only worth shipping when it's genuinely good at the task, not just technically possible.
The badge system this fits into
Every tool on Tuliq carries one of three labels, and it's meant to be checked, not taken on faith:
- 🟢 Browser — the tool never sends data anywhere, full stop.
- 🔵 Local AI — a real model, downloaded once, runs entirely on your device from then on. That's this page.
- 🟠 Server — some processing genuinely requires a server, and the tool says exactly what's sent.
If a tool's badge doesn't match what it actually does, that's the whole trust system broken — which is why it's worth being this specific about what "local AI" means rather than leaving it as a vague claim.