withoutBG

withoutBG removes image backgrounds through a managed cloud API or a free Apache-2.0 open-weight model that runs locally via Python, Docker, macOS or GIMP, with both options sharing one API schema.

withoutBG is a background removal toolkit built around two models that share one API schema. The Pro Model is a managed REST API hosted in Frankfurt, Germany, aimed at difficult subjects such as hair, fur and complex edges. The Open Model is an Apache-2.0 open-weight release that runs on the user's own hardware, with no API key and no limit on processing volume. The tool is built for developers, product teams and designers who need cutouts inside an application or workflow. Send an image to the API and get back a PNG with the background removed, or an alpha matte for compositing. Accepted API inputs include JPEG, PNG, WebP, AVIF, HEIC, TIFF, BMP and GIF, up to 20 MB per image. A browser demo on the Pro Model page allows five free images a day without signup, which makes quality checks easy before any integration work. Self-hosting comes in several forms. A Python library installs with pip or uv and processes images in a few lines, on CPU or GPU. A Docker image provides a web UI on port 80 or a headless service exposing the same /v1.0/image-without-background endpoint as the cloud. A free native macOS app handles drag-and-drop batches on Apple Silicon and can serve a local API, and a GIMP 3 plugin attaches the matte as an unapplied layer mask. ONNX weights are published on Hugging Face for custom pipelines. Privacy is a central selling point. The cloud API processes images in RAM only, discards them after the response and does not use uploads for training. Logs hold metadata, not image content. Because both options share a schema, teams can start on the API and later move to self-hosted inference without rewriting integrations. Compared with consumer cutout sites, withoutBG leans toward developers and offers documentation, datasets and comparison pages. Against other background removal APIs, its distinguishing trait is the open-weight alternative, though the open model trades some edge precision and any latency guarantee for control.