u2net-onnx

U²-Net — Salient Object Segmentation (ONNX)

ONNX checkpoints of xuebinqin/U-2-Net — a nested U-structure network for salient-object detection. Trained to separate the "main subject" of an image from the background. Pair the output mask with image_cutout() for background removal, or with apply_colormap() to visualize saliency.

Not converted locally — these are the official ONNX checkpoints, republished by danielgatis/rembg in a convenient release.

Credit: Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R. Zaiane, Martin Jagersand — University of Alberta ("U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection", Pattern Recognition 2020).

What this repo contains

Both files share the same input/output tensor signature, so inference code is identical — you can swap variants without rewriting anything.

Input / output

How to use

import onnxruntime as ort
import numpy as np
from PIL import Image

sess = ort.InferenceSession("u2netp.onnx")  # or "u2net.onnx" — same signature

# Remember the original size so we can resize the mask back at the end
orig = Image.open("photo.jpg").convert("RGB")
W, H = orig.size

# Preprocess
img = orig.resize((320, 320), Image.BILINEAR)
arr = np.asarray(img, dtype=np.float32) / 255.0
arr = (arr - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
arr = arr.transpose(2, 0, 1)[None, ...].astype(np.float32)

# Inference — outputs is a list of 7 tensors; d0 is index 0
outputs = sess.run(None, {sess.get_inputs()[0].name: arr})
d0 = outputs[0][0, 0]  # 320x320 saliency

# Normalize (U²-Net outputs aren't strictly in [0,1] before squashing)
d0 = (d0 - d0.min()) / (d0.max() - d0.min() + 1e-8)

# Resize mask back to original image dimensions
mask = Image.fromarray((d0 * 255).astype(np.uint8)).resize((W, H), Image.BILINEAR)

For background removal: apply mask as the alpha channel to the original RGB image (RGBA cutout).

Which one should I use?

• u2netp is the right default. 4.7 MB on disk, ~30 ms / image on CPU, mask quality good enough for >90% of background-removal and saliency-mapping use cases. Loads instantly.
• u2net earns its disk + latency cost on fine-edge subjects: hair, fur, lace, complex foliage, transparent objects. If the lite variant's edges look "blocky" on your inputs, the full model is the upgrade.

For interactive segmentation (clicks / boxes / prompts), pair with MobileSAM instead — U²-Net is automatic / non-interactive.

Excluded variant

The original xuebinqin/U-2-Net repo also ships a third checkpoint called u2net_portrait (line-drawing portrait sketches). It's deliberately not bundled here — it was trained on the APDrawing dataset, which carries non-commercial restrictions that would taint the otherwise-clean Apache-2.0 status of this bundle. If you need it, grab it directly from the upstream repo and read the dataset terms first.

License

Apache-2.0 — same as the upstream xuebinqin/U-2-Net repo. LICENSE file included. The danielgatis/rembg release just bundles the original weights; no relicensing occurred.