Remove background
Remove background runs a compact U²-Net in ONNX Runtime and writes a PNG with alpha. It is a product-shot helper, not a studio knockout.
← All tools
AI (on this device)
Background removal, upscale, OCR, and colorize in the browser with same-origin ONNX and Tesseract. Models aren't fetched from a CDN. Nothing is uploaded.
Runs in your browser · Nothing uploaded
- Remove backgroundCut the subject out with a small U²-Net in this tab.Runs in your browser · Nothing uploaded
- Upscale image2× or 4× in the browser. CPU wait, not an upload.Runs in your browser · Nothing uploaded
- Image to textEnglish OCR with Tesseract, worker and traineddata on this origin.Runs in your browser · Nothing uploaded
- Colorize photoGuess colour on a grayscale still, locally.Runs in your browser · Nothing uploaded
About AI (on this device)
AI on this site still has to keep the promise. Weights live in /models, ONNX Runtime wasm in /ort, Tesseract worker and English traineddata in /tesseract. The homepage doesn't prefetch any of that. You pay the download on the tool page, once, from this origin, then the service worker can cache it.
These are small models. Hair cutouts will leak. 4× upscale won't invent a lens. OCR is English print. Colorize guesses. I would rather ship that honesty than a watermarked API with five free credits.
If a future net won't fit next to a static export without a CDN, it won't ship. Subtitle generation and stem splitting are in that bucket for now, not a coming-soon paywall, just not built.
Tools in this category
Upscale image
Upscale enlarges 2× or 4×. A tiny super-resolution ONNX runs when the layout matches; otherwise high-quality bilinear, labelled as such.
Image to text
Image to text is Tesseract.js with worker, wasm, and eng.traineddata on this origin. English print. Not handwriting.
Colorize photo
Colorize tries a same-origin colourisation ONNX, then a cheap luminance tint if the weights are missing. It doesn't know the original wallpaper colour.
Common questions
Do the AI tools upload my photos to a model host?
No. connect-src doesn't allow model CDNs. Runtime files are copied into public/ at build. If the network panel shows huggingface.co while you click Remove background, that's a bug.
Why is inference slow?
WebAssembly on your CPU, single-thread, same constraint as FFmpeg here. A phone will wait. That wait is the trade for not handing the file over.
Will you add Whisper subtitles or stem splitting?
Not in this release. Those models are large or awkward to self-host without a CDN. I would rather omit them than fake a coming-soon button.
Why these tools don't upload your files
There is no server to upload to. Each page is a static file, and the work happens in your browser. Your file is read from disk into memory, processed, and handed back as a download. Once the page has loaded you can disconnect from the network and it keeps working. The longer explanation