rembg
Removes image backgrounds from the command line, a Python library, an HTTP server or Docker, using local ONNX models.
A background remover you run as a CLI, Python call, HTTP API or Docker container.
rembg strips the background from an image and hands back a transparent PNG. It suits developers and photo or product teams who want to batch-process folders or call a cutout API from their own code, not click through a web page.
What it is
You install it with pip, then run `rembg i` on one file, `rembg p` on a whole folder (with an optional watch mode), or `rembg s` to start an HTTP server. It ships many local ONNX models (u2net, isnet, BiRefNet variants, SAM, bria-rmbg) and four edge modes for hair and fur: naive, colour decontamination, alpha matting and ViTMatte.
Who it's for
Good for developers adding cutouts to an app, and for anyone with hundreds of product shots to process in one go. It is a poor fit if you want a point-and-click editor, and the default model carries a commercial-use catch you need to read first.
Key features
Four edge-refinement modes
Choose naive, -dc decontamination, -a alpha matting or -vm ViTMatte depending on how much hair or fur detail matters.
Folder watch mode
`rembg p -w` processes new or changed files in a folder as they appear.
Built-in HTTP server
`rembg s` exposes a /api/remove endpoint that takes an upload or an image URL, plus a Gradio UI you can switch off.
FFmpeg video pipe
`rembg b` reads raw RGB24 frames from FFmpeg so you can cut out video frame by frame.
Many swappable models
Pick from u2net, isnet-anime, BiRefNet portrait or general, SAM with click prompts, cloth segmentation and more.
Cloud model backends
Optionally route a job to the withoutBG or Atlas Cloud APIs from the same CLI or Python session.
Up and running in 3 steps
Our honest verdict
What's great
- One tool covers single files, folders, a Python call, a REST server and a Docker image, so the same cutout code works in a script or a service.
- SAM support lets you click a point on the subject and cut out only that object, which most one-click background removers cannot do.
- Dedicated anime, cloth and human-segmentation models exist, so you are not stuck with one general model for every image type.
- Colour decontamination removes the green or blue rim left on hair, and the README says it is cheap enough to leave on for a whole batch.
Where it falls short
- The default model, bria-rmbg 2.0, is released under a BRIA licence that the README says requires a paid agreement for commercial use, so you must switch models or pay before using it in a business.
- That default model is about 1.02 GB and runs at 1024x1024, so first run is a large download and processing is slower than with u2net.
- Python must be 3.11 to 3.13, and GPU use needs onnxruntime-gpu plus matching CUDA and cudnn, which the README flags as a common snag.
- There is no graphical editor for touch-ups; fixing a bad edge means re-running with another model or edge flag.
Pricing & license
rembg is MIT-licensed and costs nothing to run yourself. The optional withoutBG backend offers 50 free credits on signup per the README, and Atlas Cloud is a separate paid API; check each provider for rates. The bria-rmbg model needs a paid BRIA agreement for commercial use.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I use rembg for commercial work?
The code is MIT. The default bria-rmbg model is under a BRIA licence that, per the README, requires a paid agreement for commercial use, and every model carries its own licence. Check the model source before using it commercially.
Do I need a GPU?
No. `pip install "rembg[cpu]"` runs on CPU. NVIDIA (CUDA) and AMD (ROCm) GPU installs are supported for speed, but they need extra drivers and libraries.
How do I deal with messy hair edges?
The README suggests -dc first for colour halos, -a if the mask shape is wrong, and -vm for the most detail at the cost of an extra download and slower runs.
