U²-Net Human Segmentation (u2net_human_seg) background removal#

How to run u2net_human_seg

u2net_human_seg ships with rembg as the session u2net_human_seg. The weights download on first use.

pip install "rembg[cpu,cli]"
rembg i -m u2net_human_seg input.jpg output.png

from rembg import new_session, remove
from PIL import Image

session = new_session("u2net_human_seg")  # load once, reuse for every image
output = remove(Image.open("input.jpg"), session=session)
output.save("output.png")

Other sessions: all rembg models.

U²-Net Human Segmentation as a free, open-source remove.bg alternative

remove.bg is a paid cloud service. U²-Net Human Segmentation is open weights under Apache-2.0, so commercial use is allowed. It runs on your own machine (rembg i -m u2net_human_seg input.jpg output.png), offline, with no upload and no per-image cost. On Bench v0 it ranks #15 of 17 by accuracy (MAE); withoutBG Open Weights ranks #3.

remove.bg is not in Bench v0, so there is no score against it. See withoutBG Open Weights and remove.bg side by side, or compare all remove.bg alternatives.

Quick take

• Best for: Photos where the subject is a person
• Avoid when: Products, animals or objects: it is trained to find people only
• Notes: Bench v0 is general-purpose photos, so its score here understates this model on people-only images; check the People examples.

When a hosted model is worth it

u2net_human_seg is free and runs on your machine; keep it if the results above are good enough. If you need finer hair and edges, or don't want to run large models locally, rembg also ships the withoutBG Pro Model as the withoutbg session, ranked #1 of 17 by accuracy on Bench v0. Images are sent to withoutBG's servers.

export WITHOUTBG_API_KEY=sk_...
rembg i -m withoutbg input.jpg output.png

Sign up for 50 free credits (no credit card) or See its results.