# Remove an Image Background — On Your Own Device > Cut the subject out of a photo with a real segmentation model that runs in your browser. No upload, no sign-up, no watermark. This background remover runs a real segmentation model entirely in your browser — RMBG-1.4 for any subject, or MODNet, a 7 MB portrait matting model, when the subject is a person. The cut-out edge can be feathered, and shifted inward to remove the rim of old background colour that segmentation often leaves behind. The model is cached after the first run and every image is processed on your own device, so the photo is never uploaded, there is no sign-up, no per-image limit and no watermark. Output can be transparent PNG, a solid colour, or the original blurred behind the subject. **URL:** https://convertto.tech/t/ai-background-remover **Category:** On-Device AI (https://convertto.tech/c/local-ai-tools) **Privacy:** Runs entirely in the browser; no upload **Cost:** Free, no sign-up **Last updated:** 2026-08-01 ## Key facts - **Runs on:** WebGPU where available, WebAssembly otherwise - **Uploads:** None — the model is downloaded to you, not the reverse - **General model:** RMBG-1.4 — about 42 MB, downloaded once - **Fast model:** MODNet — about 7 MB, portraits only - **Maximum quality:** BiRefNet-lite on a capable GPU (~109 MB), otherwise full-precision RMBG-1.4 (~168 MB) - **Limit:** None. Unlimited images, no account - **Privacy:** Runs entirely in your browser — nothing is uploaded - **Cost:** Free, unlimited, no sign-up ## How to use 1. Select your image — the file stays on your device and is never uploaded. 2. Choose the model. 3. Choose the background. 4. Pick the colour. 5. Set the edge softness. 6. Set the edge shift. 7. Press Run, then download the result when it is ready. ## FAQ ### Is my photo uploaded? No. The model is downloaded to your browser and the segmentation runs on your device. You can disconnect from the network after the first run and it still works — which is not something an upload-based service can offer at any price. ### Why is the first run slow? The weights download once — about 42 MB for the general model, or 7 MB for the fast one. They are then cached by the browser, so every later image starts immediately. If your browser exposes WebGPU the inference takes a second or two; on the WebAssembly fallback expect ten to thirty. ### What does "needs a capable GPU" mean? BiRefNet binds 17 storage buffers per shader stage where a lot of hardware allows 16, and it does not fit the WebAssembly heap either. The tool first asks your GPU for its own maximum limits rather than accepting the low defaults — which is often enough on its own — and only if that still falls short does it use full-precision RMBG-1.4 instead. Either way the result names the model that produced it. ### There is a faint rim of the old background around my subject. That is the mask landing a pixel wide, and it is what "Edge shift" is for. Pull it to around −1 and the edge moves inward, taking the fringe with it. A little edge softness on top hides the resulting stair-step. Doing it in that order — shift first, then feather — is what a compositor would do. ### Which model should I pick? The general model unless the subject is a person and you want the smallest possible download. MODNet is portrait matting — trained to find a human — so it is outstanding on hair and edges but will not reliably isolate a product, a piece of furniture or an animal. RMBG-1.4 handles all of those. ### Does it need a GPU? No. It uses WebGPU where your browser exposes it, because that is several times faster, and falls back to WebAssembly on the CPU otherwise. The output is identical either way — only the wait changes. Models that a particular GPU cannot compile are detected before anything downloads and run on the CPU instead. ### How does it compare with the paid services? Honestly: on hair, fur and semi-transparent edges the large commercial models are still better, because they are much bigger than anything that fits in a browser. On ordinary subjects against a distinguishable background the difference is small — and this one has no limit, no account and no upload. ## Sources - [RMBG-1.4 model card](https://huggingface.co/briaai/RMBG-1.4) — BRIA AI - [MODNet: Trimap-Free Portrait Matting in Real Time](https://arxiv.org/abs/2011.11961) — arXiv - [WebGPU specification](https://www.w3.org/TR/webgpu/) — W3C ## Related tools - [AI Alt Text Generator](https://convertto.tech/t/ai-alt-text-generator): Describe an image with a captioning model running in your browser — a first draft for the alt attribute you then edit. - [AI Audio Transcription](https://convertto.tech/t/ai-audio-transcription): Run Whisper locally to turn speech into text, SRT, VTT or a timestamped transcript. Works offline after the first run. - [AI Text to Speech](https://convertto.tech/t/ai-text-to-speech): Turn text into natural speech with a neural voice running in your browser, and download the audio as a WAV file. - [Chat with PDF](https://convertto.tech/t/chat-with-pdf): Ask questions about any PDF and get answers with the page each one came from. The model runs in your browser — the file is never uploaded. - [Room Design Visualizer](https://convertto.tech/t/room-design-visualizer): Click any wall, floor or piece of furniture in your own photo, then repaint it, remove it, or swap in a different product — all in your browser. - [Aspect Ratio Calculator](https://convertto.tech/t/aspect-ratio-calculator): Work out the aspect ratio of any dimensions and scale them proportionally. - [AVIF Converter](https://convertto.tech/t/avif-converter): Convert images to and from AVIF in your browser, with a real encoder check so you never get a PNG renamed to .avif. - [EXIF Remover](https://convertto.tech/t/exif-remover): Strip all metadata — including GPS location — from photos before sharing them.