rembg

If you have a modified version of a model (e.g. converted to a different ONNX IR version for compatibility with an older CUDA toolkit), you can prevent rembg from overwriting it by following these steps:

1. Set MODEL_CHECKSUM_DISABLED=1.
2. Place your custom .onnx file in the models directory (default ~/.u2net/) with the expected filename (e.g. u2net.onnx).
3. Rembg will detect the existing file and use it instead of re-downloading the original.

export MODEL_CHECKSUM_DISABLED=1
# Place custom u2net.onnx in ~/.u2net/
rembg i input.png

To remove a background from an image, use the remove() function. By default, it uses the u2net model. You must provide a PIL Image object as the input.

from PIL import Image
from rembg import remove

input_path = 'input.png'
output_path = 'output.png'

input_img = Image.open(input_path)
output = remove(input_img)
output.save(output_path)

The withoutbg model is a cloud API backend. Images sent to this backend have a maximum upload size of 20 MB. To use it, you must provide an API key via the -x CLI flag, the api_key parameter in new_session(...), or by setting the WITHOUTBG_API_KEY environment variable.

# Example using environment variable
export WITHOUTBG_API_KEY='your_api_key_here'
rembg i input.png

Calling remove() without a session causes a new session to be initialized every time, which is a significant performance bottleneck. For processing multiple images, initialize a single session once and reuse it for every call.

model_name = "unet"
rembg_session = new_session(model_name)
for img in images:
    output = remove(img, session=rembg_session)

Install rembg using pip. Choose the backend that matches your hardware:

CPU

pip install "rembg[cpu,cli]" # for library + cli

NVIDIA GPU (CUDA)

Ensure your system is compatible with onnxruntime-gpu. NVIDIA GPUs may require onnxruntime-gpu, CUDA, and cudnn-devel.

pip install "rembg[gpu,cli]" # for library + cli

AMD GPU (ROCm)

Requires onnxruntime-rocm to be installed following AMD's documentation.

pip install "rembg[rocm,cli]" # for library + cli

Requirements: Python >=3.11, <3.14.

Run rembg in a containerized environment.

CPU Only

docker run -v .:/data danielgatis/rembg i /data/input.png /data/output.png

NVIDIA CUDA GPU

Requires the NVIDIA Container Toolkit. Because CUDA acceleration requires cudnn-devel, you must build the image locally.

1. Build: docker build -t rembg-nvidia-cuda-cudnn-gpu -f Dockerfile_nvidia_cuda_cudnn_gpu .
2. Run: sudo docker run --rm -it --gpus all -v /dev/dri:/dev/dri -v $PWD:/data rembg-nvidia-cuda-cudnn-gpu i -m birefnet-general /data/input.png /data/output.png

Tip: Use -v /path/to/models/:/root/.u2net to persist model files and avoid re-downloads.

To use the rembg command-line interface, you must install the package with the cli extra. Depending on your hardware, use one of the following commands:

• For CPU support: pip install "rembg[cpu,cli]"
• For NVIDIA/CUDA GPU support: pip install "rembg[gpu,cli]"

pip install "rembg[cpu,cli]"  # for CPU
pip install "rembg[gpu,cli]"  # for NVIDIA/CUDA GPU

To install rembg with support for both the Command Line Interface (CLI) and GPU acceleration, use the following pip command:

pip install "rembg[gpu,cli]"

! pip install "rembg[gpu,cli]"

Use the bgcolor argument to replace the removed background with a specific color. The argument accepts a tuple representing the color (e.g., RGBA).

# Example: Replace background with white
output = remove(input_img, bgcolor=(255, 255, 255, 255))

Use the post_process_mask argument to apply post-processing to the mask to achieve better results.

output = remove(input_img, post_process_mask=True)

If you do not want the original image with a transparent background, but instead want only the binary mask, set only_mask=True.

output = remove(input_img, only_mask=True)

To use a model other than the default u2net, use new_session(model_name) to create a session and pass it to the remove() function via the session argument.

from rembg import new_session, remove

model_name = "isnet-general-use"
session = new_session(model_name)
output = remove(input_img, session=session)