Install rembg 2.0.84: Remove Background AI Setup

Deploy rembg 2.0.84 for remove background AI locally. Exact pip/Docker commands, Python 3.11+ specs, model path fixes, and install errors that block first runs.

Riley Brooks2026-09-176 min readIntermediate

Why does every “simple” remove background AI install stall on the first real image? I hit that wall deploying rembg 2.0.84 – wrong Python, a ~1 GB default model, and a GPU that pretended to work. This walkthrough gets you from zero to a verified cutout without the outdated u2net-only scripts still floating around.

rembg is the open-source local tool (CLI, library, HTTP server, Docker) from danielgatis/rembg. Offline after models cache. Code is MIT; some weight files are not. Confirm the 2.0.84 tag on GitHub Releases and PyPI before you pin – do not trust mirrored “latest” blog numbers.

System requirements before you touch pip

Python gate first. Official README: >= 3.11 and < 3.14. onnxruntime is why that band exists.

python3 --version
# expect 3.11.x, 3.12.x, or 3.13.x

Windows, macOS, Linux all work. What the default model actually forces on a desk (operator judgment, not a vendor matrix):

Resource	Minimum to install	Comfortable for default model
RAM	4 GB (tiny models only)	8-16 GB (bria-rmbg ~1.02 GB on disk, heavier resident)
Disk free	~500 MB package + onnxruntime	3+ GB if you keep several ONNX files
GPU (optional)	None – CPU path works	NVIDIA + matching CUDA/cuDNN for onnxruntime-gpu, or AMD ROCm

Prove the pipeline on CPU before you touch GPU extras. Most “GPU install failed” threads are CUDA/cuDNN skew, not rembg itself.

Official download sources (use these only)

Source / docs:https://github.com/danielgatis/rembg
Python package:https://pypi.org/project/rembg/ – pin 2.0.84 after you verify the tag
Docker:danielgatis/rembg on Docker Hub (tag 2.0.84 or a digest you trust)
Release notes: same GitHub Releases page for the tag you install

Pin in production. A naked pip install -U will change behavior under you.

Install rembg 2.0.84 step by step

Virtualenv. Always. Global site-packages and distro Python fight you later.

python3 -m venv ~/.venvs/rembg
source ~/.venvs/rembg/bin/activate # Windows: .venvScriptsactivate
python -m pip install -U pip

# CPU + CLI (recommended first path)
python -m pip install "rembg[cpu,cli]==2.0.84"

# NVIDIA GPU path (only after checking onnxruntime.ai install matrix)
# python -m pip install "rembg[gpu,cli]==2.0.84"

# AMD ROCm: install onnxruntime-rocm per AMD docs first, then:
# python -m pip install "rembg[rocm,cli]==2.0.84"

Quotes around extras matter on zsh. Library-only? Drop cli – rembg[cpu] – if you only import.

Pro tip: Prefetch before a demo or CI job: rembg d bria-rmbg. First inference otherwise blocks on a ~1.02 GB fetch into the model cache.

Docker CPU (no local Python):

docker pull danielgatis/rembg:2.0.84
docker run --rm -v "$PWD":/data danielgatis/rembg:2.0.84 i /data/input.png /data/output.png

GPU image is not a Hub one-liner. Build Dockerfile_nvidia_cuda_cudnn_gpu from the repo (NVIDIA Container Toolkit; ballpark ~11 GB disk vs ~1.6 GB compressed CPU). Tag 2.0.84 added missing git in that Dockerfile. Mount -v /path/to/models:/root/.rembg so every container does not re-download weights.

First-time configuration that actually matters

No config file. Extras, env vars, and the model name you pass drive behavior.

Model home: New files land in ~/.rembg/models/<model>/ (or $REMBG_HOME). Legacy flat ~/.u2net/ is still read. Prefer REMBG_HOME.
Migrate old caches:rembg m moves legacy trees into the new layout without re-downloading when the files are intact (README models/storage section).
Default model trap: In the 2.0.8x line the default is bria-rmbg – strong edges, slower than u2net, ~1.02 GB, 1024×1024. Weights use BRIA terms: commercial use needs a paid BRIA agreement. rembg’s MIT license does not cover those weights. Need speed or clear shop rights? Force -m u2net (or another model you have actually cleared).

That license split is the line I wish the first tutorial I followed had printed in bold.

Verify the install works

rembg --version
# rembg, version 2.0.84

rembg d bria-rmbg # or: rembg d u2net
rembg i path/to/sample.jpg path/to/sample.out.png

Library smoke test:

python -c "from rembg import remove; print('ok')"

Optional server: rembg s --host 127.0.0.1 --port 7000, then http://127.0.0.1:7000/api. Add --no-ui if an optional front-end is chewing idle CPU you do not want.

On GPU builds, watch stderr on first session create. 2.0.83+ warns when you asked for CUDA (or another accelerator) and the session fell back to CPU – the old silent 10-20× slowdown (release notes / #841-class fixes).

Common install errors and fixes

“No onnxruntime backend found” on Windows after a clean rembg[cpu] install? Thread #826 pins missing Microsoft Visual C++ 2015-2022 Redistributable (x64), not a wrong extras string. Install the redist, reboot if Windows asks, then pip install "rembg[cpu,cli]==2.0.84" --force-reinstall.

CLI says CLI dependencies are not installed: library-only extras. Put ,cli back in the brackets.

GPU path installs but runs like a potato: onnxruntime-gpu wants matching CUDA/cuDNN – start at the onnxruntime install matrix, not random blog CUDA builds. Incomplete libs? Fall back to rembg[cpu,cli] instead of burning an afternoon. Remember the 2.0.83 warning above: “installed GPU” is not the same as “session using GPU.”

First run “hangs”: model download. Prefetch with rembg d .... Partial junk under the cache dir is common on flaky links; clear orphans carefully, then retry.

Python 3.10 or 3.14: outside the supported band. 3.12 is the boring safe middle.

Upgrade, migrate, uninstall

# upgrade inside the same venv
python -m pip install -U "rembg[cpu,cli]==2.0.84"

# after jumps from old caches
rembg m

# uninstall package only
python -m pip uninstall rembg

# full cleanup: package + model cache (destructive)
# rm -rf ~/.rembg ~/.u2net # Unix
# rmdir /s %USERPROFILE%.rembg & same for .u2net on Windows

Docker: pull the new tag, recreate containers. Rebuild the NVIDIA Dockerfile when the GPU image must move – Hub CPU tags do not replace that custom build.

After install sticks: rembg p for folders, rembg b for FFmpeg RGB24 pipes, or cutouts dropped into an art pipeline. Keep model name + license next to the deploy notes.

Grab one product photo. Run rembg i with an explicit -m you have licensed. Check PNG alpha before you script a hundred more.

FAQ

What is the latest rembg version I should install?

Pin ==2.0.84 after you confirm that tag on PyPI and GitHub Releases. Reproducible beats floating latest.

Do I need a GPU for remove background AI with rembg?

No. CPU extras are the path most laptops should start on. Product shots and simple portraits finish in a few seconds on a modern CPU with u2net or u2netp. Bring GPU when you batch thousands of frames or insist on heavy 1024×1024 defaults and care about latency. CUDA fight? Stay on CPU – same weights, longer wall clock.

Why is the first run huge and is the default model free for shops?

Default weights are bria-rmbg (~1.02 GB). MIT covers rembg’s code, not those weights – BRIA’s terms expect a paid agreement for commercial use. Read the model card before client catalogs. Many teams still pass -m u2net / u2netp after checking each license. Prefetch with rembg d. Files live under ~/.rembg/models/; rembg m migrates older ~/.u2net trees instead of deleting them.