Remova fundos. Direto no seu navegador.
Um removedor de fundo com IA totalmente no cliente. Sem uploads, sem servidores, sem cadastro — suas fotos nunca saem do seu dispositivo.
Privado por design. Rápido por padrão.
Sem contas, sem chaves de API, sem limites de uso. Apenas um navegador e um modelo.
100% privado
Suas imagens nunca saem do seu dispositivo. Sem uploads, sem servidores, sem rastreamento.
Acelerado por WebGPU
Roda na sua GPU quando disponível. Volta automaticamente para o WASM.
IA on-device
Impulsionado por um modelo de IA on-device via Transformers.js. Mattes com qualidade de estúdio.
Bordas perfeitas ao pixel
Detalhe fio a fio de cabelo e alpha mattes suaves — mesmo em sujeitos complexos.
Três passos. Zero servidores.
- 01
Envie uma imagem
Arraste e solte, cole ou navegue. JPG, PNG ou WebP — até o limite da memória do seu dispositivo.
- 02
A IA remove o fundo
O modelo de IA roda localmente via WebGPU e gera uma alpha matte precisa em segundos.
- 03
Baixe o recorte
Obtenha um PNG transparente na hora. Ou componha sobre uma cor sólida com um clique.
About the background remover
A free AI background remover that produces transparent PNGs, white-background product shots, and social-ready crops entirely on your device. It is built for product photos, profile pictures, portraits, and design mock-ups where uploading client or personal images to a cloud editor is not an option — and for anyone who refuses a per-image fee for something their own GPU can do in seconds.
Three quality tiers, a manual edge brush, and a batch queue with one-click ZIP export make it usable for a single avatar as well as for a two-hundred-image storefront refresh.
Three models, one decision
Fast runs MODNet (~26 MB), a portrait matting network that returns a soft alpha matte in well under a second on WebGPU — ideal for people, avatars and quick previews. Balanced and High quality both run BiRefNet Lite (~183 MB, MIT-licensed), a general-purpose dichotomous segmentation model that handles products, animals, logos and awkward edges far better than portrait-only matting. Balanced pins inference to CPU/WebAssembly so it works in every browser; High quality uses the same weights through WebGPU and is the sharpest — and slowest — of the three. Every model runs locally through Transformers.js, and the weights are downloaded once then cached by the browser.
Refining edges the model got wrong
No matting model is perfect: stray hair, a chipped mug handle, a translucent veil, or a patch of background that shares the subject’s colour. The edge brush edits the alpha matte itself rather than the pixels. Erase wipes alpha to zero; Restore writes back the exact matte value the model produced, so semi-transparent hair strands return at their original opacity instead of a hard 100 % edge. Strokes are recorded per gesture with full undo and redo, and a 3× magnifier follows the cursor for hair-level work.
Three sliders cover the rest: Feather softens the cutout boundary, Shrink / Grow removes the halo or recovers thin hair the model dropped, and Edge contrast makes the transition crisper or more gradual. Brush strokes and sliders live in separate layers, so moving a slider never destroys manual work.
From cutout to deliverable
The export panel composites the cutout onto a transparent, solid or gradient background, adds a drop shadow, fits it into the aspect ratio a marketplace or social network expects (1:1, 4:5, 9:16, 1.91:1 and more), and resizes it to a native, 1K, 2K or 4K long edge. Output is PNG when you need transparency, JPEG for the smallest white-background product shot, or WebP for both. A “trim transparent edges” switch tightens the canvas around the subject before compositing. The on-screen preview is deliberately rendered at screen resolution; the full-resolution image is only composed when you export.
Batch mode
Drop up to 500 images and they are processed one at a time in the background — one at a time because a single ONNX session already saturates the WebAssembly heap, and parallel inference is the fastest way to crash a browser tab. Each finished image is encoded to a transparent PNG immediately, so the queue never holds more than a handful of full-resolution bitmaps in memory. When the run finishes, download everything as one ZIP. Changing the export settings afterwards only re-packs the ZIP: the model never runs twice.
Limits & requirements
Chrome or Edge with WebGPU is fastest; everywhere else the models fall back to WebAssembly with identical quality at a slower pace. Inference runs at a fixed internal resolution (512×512 for the BiRefNet Lite graph, 512 for MODNet), so the matte is upscaled to your photo’s full resolution on export — a 5000px photo and a 1600px photo therefore carry exactly the same model detail. That is why the tool caps the image it feeds the network, and why huge files do not get sharper. Camouflaged subjects (a snowman on snow, a black cat on a black sofa) may still need the brush. There is no file-size limit and no watermark.
Privacy
Images are decoded, processed and encoded in your browser’s memory; nothing is uploaded, logged, or stored on a server. There is no account, no usage counter, and no telemetry attached to your files. As a side effect, the tool keeps working offline once the model weights are cached.