Web tools like remove.bg are fine for one image, but they cap free downloads at low resolution and every file leaves your machine. rembg does the same job locally, at full resolution, on as many files as you like. It runs a segmentation model that predicts a per-pixel alpha mask, then writes a transparent PNG — plain local inference, no account.
1. Install
pip install "rembg[cpu,cli]"
Installs the library plus the rembg command. This has changed: a bare pip install rembg no longer pulls a runtime — you must pick an extra. Use "rembg[gpu,cli]" for NVIDIA/CUDA or "rembg[rocm,cli]" for AMD. Python 3.11, 3.12 or 3.13 is required.
2. Cut one image
rembg i input.jpg output.png
Reads one file, writes a transparent PNG beside it. The first run downloads the model, so it is slow once and fast after.
3. Batch a folder
rembg p ./photos ./cutouts
Processes every image in ./photos into ./cutouts, reusing the loaded model instead of paying startup cost per file. There is also rembg s --host 0.0.0.0 --port 7000 if you would rather call it as a local HTTP service.
4. Change the model when edges are rough
rembg i -m isnet-general-use input.jpg output.png
Swaps the segmentation model. The default is bria-rmbg; isnet-general-use and the birefnet-* family often hold hair and thin objects better, u2net_human_seg is tuned for people, and isnet-anime for illustration.
Verify it worked
python -c "from PIL import Image; print(Image.open('output.png').mode)"
Prints the colour mode — you want RGBA. RGB means the alpha channel was dropped, almost always because the output filename ended in .jpg, which cannot store transparency.
If it fails: command not found means you installed without the cli extra. Fuzzy or missing edges usually mean low contrast between subject and background rather than a bad model — reshoot or try -m birefnet-general. For a mask you want to paint by hand instead, see inpainting; to enlarge the cutout afterwards, upscale it locally. Model list and options: the rembg README.