withoutBG: Open-Weights Background Removal Model
Apache-2.0 · Python · Docker · macOS · GIMP
How to run it#
Self-host with a Mac app, GIMP plugin, Python package, or Docker. Same open weights across the stack. See also self-hosted background removal.
macOS App
- Native desktop app with drag-and-drop batch processing
- Optional Local API for GIMP, Blender, and scripts on localhost
- On-device Apple Silicon inference, works offline after install
GIMP Plugin
- Remove backgrounds from the GIMP 3.0 Tools menu
- Calls a local withoutBG server on localhost
- Attaches the alpha matte as an unapplied layer mask
Python Library
- Install with uv add withoutbg
- Process images in 3 lines of Python
- Runs on CPU or GPU, Apache-2.0 licensed
Docker
- Web UI: one docker run command, inference on port 80
- Headless API: same /v1.0/image-without-background schema as cloud
- Images stay on your network, nothing leaves the container
New in v4: Smart Brush#
Missed a strand, or kept a chair you did not want? Paint Keep over what went missing and Remove over what should go. A small stroke-edit network (2.5 MB, open weights like the rest) fixes the cutout. One stroke on an object is enough; pixels away from your strokes do not change. It works in the Mac app (Smart Brush), the Docker web app, the Python package (model.edit(image, keep=..., remove=..., matte=...)) and the Hugging Face Space. On our frozen validation set, one round of strokes cuts the error inside the edited region by 83% on matte fixes and 96% on unwanted or missing objects. More about Smart Brush.
Coming next: one gateway, two specialist models#
The next community release automatically routes each image to the model suited to its edges. withoutBG trains and maintains the matting branch for hair, fur, soft detail, and transparency. BiRefNet handles segmentation for opaque objects, flat scenes, and vehicles.
Matting → withoutBG
Fine strands, soft boundaries, and partially transparent regions.
Segmentation → BiRefNet
Solid subjects and clearly defined object boundaries.
This pipeline is being prepared for Python, the Mac app, Docker, and Hugging Face. GIMP uses the gateway in its connected Mac or Docker server. Routing runs wherever inference runs; local installations keep images on your device. Existing downloads and benchmark results still refer to the released model until the new bundles ship.
Examples#
Sample outputs from the open-weights model. Switch between transparency, chroma key, and alpha matte on each result.
Photo by Niko Tsviliov on Unsplash
Compare to other open models#
Same Bench images and ground-truth metrics against rembg sessions (u2net, IS-Net, BiRefNet, and more). The directory is the evaluation surface; this page is where you install and ship.
WithoutBG Ecosystem
ONNX weights on Hugging Face and open-source repositories.
Community packages
Built by others on the open weights. Not maintained by withoutBG.
Get notified about new releases#
New weights, versions and integrations, by email. No account needed, and every email has a one-click unsubscribe.