About withoutBG#

withoutBG builds background removal models. The open-weights model runs on your own hardware. The API model runs as a hosted service. Operating since 2022, based in Berlin.

Measured, not claimed#

Every model is tested against paid services and open-source alternatives on the same images.
The results are published, including the cases where withoutBG scores worse.
A quality claim on this site should come with a number you can check.

The open-weights model on the public benchmark set, against widely used open models:

Open weights vs	MAE	Gradient error
BiRefNet General	0.013 vs 0.017 better	0.035 vs 0.034 worse
BiRefNet Lite	0.013 vs 0.030 better	0.035 vs 0.035 better
IS-Net	0.013 vs 0.041 better	0.035 vs 0.046 better
u2net	0.013 vs 0.067 better	0.035 vs 0.054 better
silueta	0.013 vs 0.071 better	0.035 vs 0.058 better
u2netp	0.013 vs 0.072 better	0.035 vs 0.055 better

Mean error against ground-truth alpha mattes on Bench v0 (33 images, same inputs for every model). Lower is better. Each row links to the side-by-side images.

The rows that say “worse” stay in the table. Side-by-side results for the API model, including against remove.bg, are in Compare.

How results are measured#

Same input images for every model (benchmark methodology).
Standard alpha matting metrics: SAD, MSE and gradient error (what each metric measures).
A public evaluation set: withoutBG100.
300+ tracked training experiments. A model change ships when the numbers improve.

Data, training and evaluation in detail: Engineering background removal.

Products#

API: REST endpoints for background removal and alpha mattes.
Web app: batch background removal in the browser, using the API model.
Open weights: Python, Docker, CLI and ONNX on Hugging Face.
Mac app: the open-weights model, offline on your Mac.
Figma plugin and Shopify app.
Car API: cutouts, shadows and studio backgrounds for vehicle photos.

Hosting and privacy#

The API runs on AWS in Frankfurt. An EU region was chosen for GDPR: most customers are in Europe.
Images sent to the API are not stored (privacy policy).
In the web app, the original image stays on your device. Only a reduced copy is sent to compute the mask.
The open-weights model sends nothing anywhere.

Licensing#

Open weights: Apache 2.0 for the withoutBG parts. The DINOv3 backbone is under the Meta DINOv3 License. Details: license.
API model: closed weights, paid per image with credits (pricing).

Company#

Operating since 2022 in Berlin, founded by Imran Kocabiyik. Legal details: imprint.
Support in English: contact form or contact@withoutbg.com.
Code and models: GitHub, Hugging Face, Docker Hub, PyPI.