AI tools that run on your own machine

What these four tools do

AI Image Maker writes a picture from a line of text. Photo AI Studio does the rest of the picture work in one place: it cuts a background out, blurs one, enlarges a small image, describes a photo for alt text, reads any writing in it, ranks stock keywords and finds duplicates — and it can paint a new picture too. Image to Text reads writing out of a photograph. Writing Studio proofreads and summarises.

Background Remover & Photo AI lifts a subject away from its background, the job behind most product shots and headshots. Image Enlarge adds pixels to a small picture. People reach for this category when the task is small and personal: one photo, one recording, one paragraph, and no wish to open an account for it.

Why a browser is an odd place for this

Usually your file travels to the model. Here it is the other way round. The first run fetches a model file, and your browser keeps it. From then on the picture, the recording or the text stays with you. That reversal is the whole design, and it is what makes the promise of nothing uploaded hold for AI work.

The trade-off is real. Your own machine is doing what a server normally does, so speed depends on the hardware in front of you, and these models are smaller than the ones behind paid services. Some need a recent browser with graphics support. Expect a wait on the first run, and expect good results rather than flawless ones.

Worth knowing before you start

Do the first run on a decent connection, and on the device you actually plan to use. The download is held by that browser, so a different machine, or a cleared cache, means fetching it again. Nothing is lost when that happens. It simply costs you the wait a second time.

The common mistake is with Image Enlarge. It cannot recover detail the camera never captured. It invents plausible pixels, which suits a soft photo and can smudge fine text. On the other side, nothing here charges by the run, so with AI Image Maker the sensible tactic is many quick attempts rather than one careful one.

The size of the first download, tool by tool

The download is not one size, and the range here is wide. Image to Text needs about 45 MB. Photo AI Studio fetches only what the job you picked needs: about 90 MB to cut a background out, 30 MB for portrait blur, 95 MB for keywords and duplicates, and 265 MB to describe a picture or read text from it. Writing Studio is about 570 MB. Painting a picture — in the Image Maker or in Photo AI Studio — is close to a gigabyte, nearly all of it sitting in one weights file.

Photo AI Studio splits its download across its jobs, so opening the page fetches nothing at all. Cutting out a background costs about 104 MB. Judging depth for a portrait blur costs about 27 MB, and enlarging about 18 MB. The jobs that write about a picture are the heavy ones: the model behind Describe is about 288 MB on its own, and it takes roughly half a minute per caption.

Two smaller habits are worth knowing. If your browser reports Data Saver, no tool starts fetching early, so the wait moves to the moment you press the button. And under every AI tool there is a panel showing how much AI data this browser is holding, with a button that deletes it. The figure comes from the browser's own storage estimate, so it counts your disk rather than a guess. Deleting it costs you nothing but the download next time.

The weight format was measured, not assumed

Almost all of this runs on WebAssembly, on a single thread. More threads would need shared memory, shared memory needs cross-origin isolation, and the site cannot switch that on without breaking the ads that pay for it. That one limit shapes the choices below.

Smaller weights are not always better. The keyword model has a 4-bit build at half the size, and it was tried and dropped: a desktop runtime finishes it in 60 milliseconds, while the WebAssembly build never finishes at all. The writing model went the other way. Its 8-bit build is 348 MB against 370 MB and writes about three times faster, and it was still rejected, because at 8 bits it turned "within thirty (30) days" into "within three months", and "the leak has not worsened" into "worsening leaks". Slow and honest won that argument.

A GPU does not always help either. The captioning model was timed on WebGPU and came out slower than WebAssembly, 39 seconds against 30, and the files that would suit a GPU would add about 180 MB. Unlimited AI Image Maker is the one tool that genuinely needs WebGPU. It says so plainly instead of falling back, because on WebAssembly it would take minutes per picture. Enlarge makes the same point from the other side: its model was swapped in August 2026 for one built from convolutions, which run about thirty times faster on one thread than the transformer it replaced.

Jobs that were removed, and the ones that overlap

Writing Studio started with seven jobs and ships with three. Make it shorter, plain English, make it polite and title ideas were all built, run over real text, read, and taken out. Each one failed the same way: it changed the meaning and gave no sign that it had. A complaint about a missed deadline came back apologising for having missed it. Two hours of downtime on a Sunday morning became a subject line saying the site was down until 2pm. Three different prompt shapes were tried, and none of them moved the failure, so the four jobs are gone.

Dictation was pulled once as well. An earlier version leaned on the browser's own speech feature, which sends the recording to Google. That breaks the promise this site makes on every page, so it was withdrawn and will not come back in that form. The on-device replacement built on Whisper was itself parked in September 2026 while the site was cut back, and returns with the rest.

Two jobs read the words in a picture, and they are not interchangeable. Image to Text (OCR) is built for dense typed pages: receipts, letters, forms. It scales small pictures up before reading, because that engine wants letters around 30 pixels tall and gets much worse below 20. Read text, inside Photo AI Studio, is the one for a sign, a label or a screenshot.

Describe and Keywords also look alike, and they are two different models. One writes a sentence. The other only scores your picture against a fixed list of stock words, so the first few tags are good and the tail is noise; the scores sit in a narrow band with no clean gap between a right answer and a wrong one. That is also why there is no box for finding a photo by typing a description. It would need the half of that model the site does not ship. Find duplicates works because comparing one picture with another needs nothing extra.