BackgroundRemover

BackgroundRemover is a command-line background-removal tool open-sourced by Johnathan Nader (nadermx). Built on the U^2-Net saliency detection model and a PyTorch inference stack, it runs on Linux, macOS, and Windows and pulls foreground subjects out of local images or videos, producing transparent PNG, MOV, or GIF outputs. The project also exposes a Python library entry point and a standalone HTTP API server, so it can plug into data pipelines as a backend capability.

Core capabilities span three layers. First, the model layer ships three pretrained weights — u2net (general purpose, default), u2net_human_seg (portraits and human subjects), and u2netp (lightweight, fast) — switchable via -m. Second, the image layer accepts JPG, PNG, HEIC, and HEIF inputs and supports alpha matting edge refinement (-a, -af, -ab, -ae), mask-only output, custom solid-color backgrounds, custom image compositing, and stdin/stdout pipelining for batch jobs. Third, the video layer accepts MP4, MOV, WebM, OGG, and GIF inputs and outputs ProRes 4444 transparent MOV (default prores_ks + yuva444p10le), transparent WebM VP9, transparent GIF, and Premiere matte-key files, plus compositing onto another image or video.

Use cases include e-commerce product photos with consistent backgrounds, ID-photo and avatar portrait cleanup, social-media short-video keying and background swaps, pre-vis rough keys for film and video, and batch preprocessing of image datasets. Anywhere a subject needs to be lifted from its background without paying for a SaaS, this tool works as a one-time local investment. Installation requires Python 3.6+, PyTorch (CPU or GPU), and FFmpeg 4.4+. Three deployment shapes are available: pip install, a Docker image, and a backgroundremover-server HTTP API. Model weights cache to ~/.u2net/ on first run. Code is MIT; model weights are Apache-2.0. The author also runs a commercial hosted version at BackgroundRemoverAI.com.