Background Removal API and Open-Weight Model

Two models, one kit. Self-host the open-weight model (Apache-2.0) for full data control, or use the managed API for stronger results on hair, fur, and complex edges.

Open-weight model quick start: run with Docker using docker run -p 80:80 withoutbg/app:latest, or install via Python with pip install withoutbg. Apache-2.0 license. View source on GitHub at https://github.com/withoutbg/withoutbg

The kit

Two models with one shared API schema. Pick the deployment that fits your stack, or run both.

API Model

Pro model

Managed REST API · Frankfurt, DE · p95 ≈ 1.95 s · AWS Inferentia

API model background removal

Send an image, get back a PNG with the background removed or an alpha matte. Integrate in minutes with curl or any HTTP client. New accounts receive 50 free credits.

  • Frankfurt, DE (EU). TLS encryption in transit.
  • RAM-only processing. Images discarded after the response. No disk writes, caches, or backups.
  • Customer uploads are never used for model training.

PricingDocs

Get API key

Open-Weight Model

Apache-2.0

Self-hosted · Python, Docker, macOS · images stay on your network

Python library documentation

Python Library

uv add withoutbg. Process images in 3 lines of Python. Apache-2.0 licensed, runs on CPU or GPU, supports batch processing.

View Python docs
macOS desktop background removal app

macOS App

Native macOS desktop app with drag-and-drop background removal. Inference runs locally on your Mac. No internet connection required after install.

Get macOS app
macOS headless server for batch background removal

macOS Server

Headless inference service for local app integration. Run it so the GIMP plugin can call localhost; also works with Blender add-ons and other tools you extend on your Mac. Same HTTP API as Docker. Compiled for Apple Silicon.

Get macOS server
GIMP 3 plugin for local background removal

GIMP Plugin

Remove backgrounds inside GIMP 3.0 from the Tools menu. Sends the active layer to a local withoutBG server (Docker or Mac server), then attaches the alpha matte as an unapplied layer mask.

View GIMP docs
Docker Desktop web UI for background removal

Docker Desktop

Local web UI via Docker Desktop. One docker run command starts inference on port 80. Images stay on your network; nothing leaves the container.

View Docker docs
Docker API service for background removal

Docker Service

Headless Docker image exposing /v1.0/image-without-background on port 80. Same API schema as the cloud endpoint. Drop-in replacement for local or on-prem deployments.

View Docker docs
withoutBG ONNX model on Hugging Face

Hugging Face

ONNX model weights hosted on Hugging Face. Download directly for custom inference pipelines or to inspect the model architecture.

View on Hugging Face
withoutbg-python GitHub repository

withoutbg-python

Source code for the Python client library (Apache-2.0). Includes the CLI, batch processing utilities, and integration tests.

View on GitHub
withoutbg-inference GitHub repository

withoutbg-inference

Source code for the inference server behind the Docker image (Apache-2.0). Contains the model loading, pre/post-processing pipeline, and HTTP API layer.

View on GitHub

Specs and limitations

Specifications, security, training data, and known limitations for the API and open-weight models.

Security and privacy

  • Open-weight: Run locally; no API calls required. Full control over image data.

  • API: Frankfurt, DE (EU). TLS in transit. RAM-only processing; images discarded after the response. No disk writes, caches, or backups.

  • No customer training: Uploads are never used to train or fine-tune models. Processing buffers zeroed after response.

  • Logging: Timestamp, endpoint, duration, status, request size, API key hash. No image bytes or perceptual hashes.

  • Analytics: No cookies on the web UI. Ahrefs (privacy-friendly). In-house CAPTCHA for abuse prevention.

Training data

  • Public photos: Unsplash/Pexels under permissive licenses; we created alpha mattes. withoutBG100 dataset

  • Licensed sets: Purchased image sets with explicit derivative rights.

  • Synthetic renders: Randomized lighting/camera with ground-truth mattes from the render pipeline.

  • Studio captures: Hair, translucency, shadows. Dataset guide

  • Scale: ~60K image/matte pairs after QA (2025-10-01), expanding.

Limitations

  • API input cap: Maximum 10 MB per image (JPEG, PNG, WebP, TIFF, BMP, GIF).

  • Open-weight latency: Depends on your CPU/GPU and image resolution. No fixed SLA.

  • Subjective boundaries: When foreground vs. background is ambiguous, multiple valid cutouts exist. Future release will bias toward nearest-camera subjects.

FAQ

Common questions about the API Model, open-weight model, pricing, and privacy.

Is it really free?

Open-weight model: Apache-2.0. Commercial use, modification, and redistribution permitted at no cost. API Model: freemium, 50 free credits on signup, then pay-as-you-go.

  • Open-weight: self-host with Python, Docker, or the macOS app. No API key, no outbound data.
  • API Model: credit packs for occasional use or optional monthly subscription. Cancel any time. See pricing.
  • No subscription lock-in on either option.

API Model or open-weight model?

Both share the same API schema. Pick by deployment needs, not by integration work.

  • API Model (Pro): managed REST API in Frankfurt, DE. Stronger on hair, fur, and complex edges. p95 ≈ 1.95 s server-side.
  • Open-weight (Apache-2.0): runs on your hardware. Images never leave your network. CPU or GPU. No SLA.
  • Start with the API for quick integration; switch to self-hosted at any time using the same /v1.0/image-without-background endpoint.

How fast is it?

API Model: p95 ≈ 1.95 s server-side in Frankfurt, DE on AWS Inferentia. End-to-end time adds upload, TLS handshake, queue, and download.

  • Open-weight: latency depends on your CPU/GPU and image resolution. No fixed SLA. Typical range on Apple Silicon: 0.3–2 s.
  • API: processing starts as soon as the upload completes. The demo on /api-model/remove-background shows your real measured latency per call.

Do you store or train on my images?

No. API images are processed in RAM and discarded after the response. Open-weight inference requires no API calls at all when run locally.

  • API storage: no disk writes, thumbnails, caches, or backups. RAM-only; buffer zeroed after response.
  • Region: Frankfurt, Germany (EU). TLS in transit.
  • Logs: metadata only: timestamp, endpoint, duration, status, request size, API key hash. No image bytes or perceptual hashes.
  • Training: customer uploads are never used to train or fine-tune models. See privacy policy.

How do I self-host?

Three lines of Python or one Docker command. Apache-2.0, no API key required.

  • Python: uv add withoutbg (or pip install withoutbg). See Python docs.
  • Docker: docker run -p 80:80 withoutbg/app:latest starts a local web UI. The headless service exposes the same /v1.0/image-without-background API as the cloud. See Docker docs.
  • macOS: native desktop app and headless server for Apple Silicon, with or without internet. See macOS app.
  • Model weights are hosted on Hugging Face in ONNX format.

Supported by

NVIDIA Inception Program
AWS Activate Program