withoutbg/withoutbg-openweights-v4-service-cpu repository overview

⁠withoutBG Open Weights v4 — Inference API (CPU)

Self-hosted withoutBG⁠ open weights background removal API. Runs the v4 ONNX model on CPU with a FastAPI service — no GPU required.

New in v4: Smart brush. Paint what to keep and what to remove; a small stroke-edit network fixes the cutout. Send keep_mask / remove_mask (and the previous matte) to /v1/remove-background.

⁠Quick start

docker run --rm -p 8000:8000 withoutbg/withoutbg-openweights-v4-service-cpu:latest

Health check: http://localhost:8000/health

⁠Platforms

Published for linux/amd64 and linux/arm64. Docker pulls the matching architecture automatically (Intel/AMD, Apple Silicon, AWS Graviton, etc.).

⁠API

Same input/output schema as the withoutBG Mac Local API: send a raw JPEG/PNG body (or multipart field image); response is image/png with headers X-Latency-Ms, X-Route-Category and X-Route-Pipeline (matting or birefnet). Use ?output=matte for a grayscale alpha matte instead of the default cutout.

curl -X POST \
  --data-binary @photo.jpg \
  -H "Content-Type: image/jpeg" \
  http://127.0.0.1:8000/v1/remove-background \
  -o result.png

curl -X POST "http://127.0.0.1:8000/v1/remove-background?output=matte" \
  -H "Content-Type: image/png" \
  --data-binary @photo.png \
  -o matte.png

⁠Optional: withoutBG Pro

Set -e WITHOUTBG_API_KEY=sk_your_key -e WITHOUTBG_PRO_PASSWORD=<long random password> and POST /v1/remove-background?engine=pro uses withoutBG Pro. Pro uploads the image to api.withoutbg.com (1 credit each). The open weights stay the default and never send images anywhere. Pro calls are charged to your key, so each one must send the password (X-Pro-Password; the web UI asks once). Keep a Pro-enabled server private, e.g. -p 127.0.0.1:8080:8080. Keys: withoutBG Pro⁠.

⁠Model

The withoutBG Open Weights 10.8.0 bundle (~1.5 GB, baked into the image): a trained router sends fine strands, soft detail and transparency to the withoutBG matting model, and hard opaque objects, flat scenes and vehicles to BiRefNet. Only the selected branch runs.

⁠Related images

⁠Links

• withoutBG open model⁠
• Source on GitHub⁠
• License: Apache-2.0 for the code; the model is under the withoutBG Open Weights license⁠ and includes DINOv3, Depth Anything V2 (Apache-2.0) and BiRefNet (MIT). See /v1/licenses in the running container