Features

Projects live in your browser’s own storage on the one device that made them. They do not sync. A different laptop, a different browser, or a private window sees none of it, and clearing site data destroys the lot with nothing to restore from. We cannot recover your work, because we never had a copy.

There is also no API, no CLI, and no server-side batch endpoint — nothing to automate against. If your workflow needs a thousand images processed unattended overnight, this is the wrong tool and we would rather say so here than after you have tried.

The models run on your device, which means they have to reach your device first. The fast cutout model is around 45 MB. The best-quality model is up to 192 MB — capable GPUs fetch a smaller half-precision build, but a machine without usable WebGPU fetches the full one. On a metered or slow connection that is a real cost, paid before you see a single result.

Once cached the app genuinely works with the network off. Until then it does not work at all.

On a machine with a healthy GPU, inference runs on WebGPU. Everywhere else it runs on WebAssembly, which is materially slower — minutes rather than seconds on a large image, and considerably worse on video, where the work is multiplied by the frame count. We deliberately refuse software and degraded GPU adapters instead of using them, because a silent downgrade mid-run is worse than a slow run you were told about.

Phones and any device reporting four gigabytes of memory or less are capped further still: the image is reduced to roughly 1024–1536 px on its longest side before it enters the pipeline. That is well under the 4096 px desktop ceiling. A full-resolution phone photo is not processed at full resolution on that phone — the alternative was a tab that crashed and lost the work.

Nothing in the catalogue takes a text prompt. There is no generation, no outpainting, no inventing of pixels that were never captured — every tool here transforms media you already have.

The video upscaler in particular is not an AI tool at all. It is a classical denoise, Lanczos scale, and sharpen pipeline. It makes a small clip look better at a larger size; it cannot recover detail the camera never recorded, and we would rather label it honestly than borrow the word.

Cutout quality is a model’s judgement, not a guarantee. Fine hair against a busy background, wire-frame furniture, chain-link, and motion-blurred edges are all cases where you should expect to reach for the brush and edge-refine tools rather than the download button. The fine-detail model reasons at a 512-pixel working size and its mask is scaled up from there, which is exactly where the hardest edges lose their last few pixels of precision.

Our tool taxonomy still names categories that hold nothing: Audio. The standalone AI suite and the in-editor chat assistant were retired when the product refocused on cutouts, the editors, and upscaling, and the audio work never shipped. The labels remain because the code that would repopulate them is still wired up; the tools are not there, and no page will pretend otherwise.

Only some categories have a landing page of their own. Tools in the Model-powered, Editor category are reachable from the nav and from this page, but there is no browsable index built for them yet.

Nothing on this page is a roadmap entry. Every tool listed above is marked live in the catalogue and is usable today — if a feature is missing from this page, it is missing from the product.