Skip to content
imaged

Background removal

MODNet (portraits)

Separates a person from the background with a soft alpha matte, hair and edges included. Apache-2.0.

Apache-2.0background removalmattingportraitmodnetapache-2.0
Choose a variant

About this model

MODNet removes the background behind people. For every pixel it gives how much of it belongs to the person, from 0 (background) to 1 (person), so hair and soft edges blend cleanly into a new background instead of being cut along a hard line.

This is the photographic portrait model from the authors (6.5 million parameters, MobileNetV2 backbone). It needs no trimap or second photo of the empty background.

What you can build

  • Background replacement and blur for video calls
  • Product-style cutouts of people for photo apps
  • Stickers and portraits on a transparent background
  • Privacy: hide what is behind the person

Man with beard portrait

f16
Input
Bearded man smoking pipe-3013924

ThuyHaBich, CC0, https://commons.wikimedia.org/wiki/File:Bearded_man_smoking_pipe-3013924.jpg

Output
The person cut out by MODNet (portraits) over a checkerboard 
Matte on the Metal (FP16 pack); 43% of the photo is person (alpha above one half). Shown over a checkerboard.

People at market

f16
Input
Fishmonger smiling, Maracaibo street market, Venezuela

Wilfredor, CC0, https://commons.wikimedia.org/wiki/File:Fishmonger_smiling,_Maracaibo_street_market,_Venezuela.jpg

Output
The person cut out by MODNet (portraits) over a checkerboard 
Matte on the Metal (FP16 pack); 42% of the photo is person (alpha above one half). Shown over a checkerboard.

Basketball players

f16
Input
Basketball player 13

LTx Domas, CC0, https://commons.wikimedia.org/wiki/File:Basketball_player_13.jpg

Output
The person cut out by MODNet (portraits) over a checkerboard 
Matte on the Metal (FP16 pack); 56% of the photo is person (alpha above one half). Shown over a checkerboard.

Variants & downloads

Choose the build for your app

Model files are hosted on Google Drive. Large files may open a confirmation page before downloading.

Standard FP16Small int8

Standard

Slab

Weights stored in FP16: half the size of FP32, the same results (see Accuracy).

Platforms: macos · android

modnet-f16.slab

15.1 MB · Hosted download

Download
SHA-256 checksum
fd2be360a831843106d4cc73d311899009b8c94a7032c2ee4745a3332e06f4d9

LICENSE-MODNet.txt

11 KB

Download
SHA-256 checksum
c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4
Conversion details
command
git clone https://github.com/ZHKKKe/MODNet && git -C MODNet checkout 28165a451e4610c9d77cfdf925a94610bb2810fb # run the export next to MODNet/, not inside it: its onnx/ folder hides the onnx package gdown --folder https://drive.google.com/drive/folders/1umYmlCulvIFNaqPjwod1SayFmSRHziyR -O MODNet/pretrained # the README's Google Drive folder (pip install gdown) python -c "import sys, torch, importlib.util as u; sys.path.append('MODNet'); s = u.spec_from_file_location('m', 'MODNet/onnx/modnet_onnx.py'); m = u.module_from_spec(s); s.loader.exec_module(m); net = m.MODNet(backbone_pretrained=False); sd = torch.load('MODNet/pretrained/modnet_photographic_portrait_matting.ckpt', map_location='cpu', weights_only=True); net.load_state_dict({k[7:]: v for k, v in sd.items()}); torch.onnx.export(net.eval(), torch.randn(1, 3, 512, 512), 'model.onnx', input_names=['input'], output_names=['output'], opset_version=17, dynamo=False)" slab-load convert model.onnx model.slab --input input=1x3x512x512 --meta task=segment --meta resize=stretch --meta antialias:bool=true --meta normalize=neg_one_to_one --meta color=rgb slab-load quantize model.slab modnet-f16.slab --f16 slab-load meta set modnet-f16.slab default_backend.macos=gpu

Integration examples

Python

python

import slab

cfg = slab.Config(cache_dir="/tmp/slab_cache", license=open("license.slablic.txt").read())
session = slab.Session(cfg, "modnet-f16.slab", warmup=True)
photo = slab.Frame.open("photo.jpg")
out = session.run(slab.Input(frame=photo))
alpha = out.matte()               # Gray8 Frame, the photo's size
cutout = photo.with_alpha(alpha)  # RGBA8 Frame, background transparent
print(cutout.width, cutout.height)

Device & backend compatibility

Pack default: macos / GPU. The fastest backend on the Mac (about 10 ms per frame); the Neural Engine takes 14 ms.

Available execution backends for MODNet (portraits)
PlatformEngineStatusPrecision
apple

CPU

CPU

Supported

Measured on macOS.

f32
apple

Metal

GPU

Supported

Measured on macOS. The default.

f32
apple

Apple Neural Engine

NPU

Supported

Measured on macOS. The Neural Engine computes in FP16.

f16
android

CPU

CPU

Supported

Measured on a Samsung Galaxy S23 Ultra (Snapdragon 8 Gen 2, Adreno 740), Android 17.

f32
android

Vulkan

GPU

Supported

Measured on a Samsung Galaxy S23 Ultra (Snapdragon 8 Gen 2, Adreno 740), Android 17.

f32
android

Qualcomm Hexagon NPU (QNN)

NPU

Not supported

Tried on a Samsung Galaxy S23 Ultra (Hexagon v73): Slab's QNN backend has no translation for Expand yet.

—

Performance

Results depend on the device, runtime, and test conditions. Each measurement keeps its own context.

Mac with Apple M1 Max

Slab · CPU

macOS 26.6.2

141.0 ms

Measured time · scope below

f16Session open (ms): 12 Platform: macos Runs: 100 Input: 1, 3, 512, 512 Measured: 2026-10-28

Median of 100 warm runs of the model (Slab's run time). From a 1920 x 1080 photo, resizing included: 140.43 ms median. matte() at 1920 x 1080 adds about 3 ms.

Mac with Apple M1 Max

Slab · GPU

macOS 26.6.2

10.4 ms

Measured time · scope below

f16Session open (ms): 121 Platform: macos Runs: 100 Input: 1, 3, 512, 512 Measured: 2026-10-28

Median of 100 warm runs of the model (Slab's run time). From a 1920 x 1080 photo, resizing included: 12.32 ms median. matte() at 1920 x 1080 adds about 3 ms.

Mac with Apple M1 Max

Slab · NPU

macOS 26.6.2

13.8 ms

Measured time · scope below

f16Session open (ms): 2905 Platform: macos Runs: 100 Input: 1, 3, 512, 512 Measured: 2026-10-28

Median of 100 warm runs of the model (Slab's run time). From a 1920 x 1080 photo, resizing included: 15.67 ms median. matte() at 1920 x 1080 adds about 3 ms.

Samsung Galaxy S23 Ultra (Snapdragon 8 Gen 2)

Slab · CPU

Android 17

631.0 ms

Measured time · scope below

f16Session open (ms): 67 Platform: android Runs: 19 Input: 1, 3, 512, 512 Measured: 2026-11-06

Median of 19 warm runs of the model (Slab's run time). From a 1920 x 1080 photo, resizing included: 614 ms median. open_ms: opening the session. Peak memory (resident) after the runs: 393 MB.

Samsung Galaxy S23 Ultra (Snapdragon 8 Gen 2)

Slab · GPU

Android 17

143.0 ms

Measured time · scope below

f16Session open (ms): 93 Platform: android Runs: 68 Input: 1, 3, 512, 512 Measured: 2026-11-06

Median of 68 warm runs of the model (Slab's run time). From a 1920 x 1080 photo, resizing included: 146 ms median. open_ms: opening the session with the GPU programs already in the cache folder (the first open on a device compiles them, about 0.3 s more). Peak memory (resident) after the runs: 102 MB.

Accuracy checks

macos · CPU · f16

The upstream PyTorch model, 41 public-domain and CC0 photos of people (Wikimedia Commons). Same size: PyTorch on the same 512 x 512 stretch, its matte resized back bilinearly. Upstream script: MODNet's own inference (aspect kept, about 512 on the short side). Differences are mean |alpha| on a 0..1 scale

With Slab preprocessing and matte(): 0.19% mean difference to the same-size PyTorch matte (1.10% on the worst photo). On identical input tensors: 0.062%. Against MODNet's own script, which keeps the aspect ratio: 6.5% (25.5% on the worst photo).

macos · GPU · f16

The upstream PyTorch model, 41 public-domain and CC0 photos of people (Wikimedia Commons). Same size: PyTorch on the same 512 x 512 stretch, its matte resized back bilinearly. Upstream script: MODNet's own inference (aspect kept, about 512 on the short side). Differences are mean |alpha| on a 0..1 scale

With Slab preprocessing and matte(): 0.19% mean difference to the same-size PyTorch matte (1.10% on the worst photo). On identical input tensors: 0.062%. Against MODNet's own script, which keeps the aspect ratio: 6.5% (25.5% on the worst photo).

macos · NPU · f16

The upstream PyTorch model, 41 public-domain and CC0 photos of people (Wikimedia Commons). Same size: PyTorch on the same 512 x 512 stretch, its matte resized back bilinearly. Upstream script: MODNet's own inference (aspect kept, about 512 on the short side). Differences are mean |alpha| on a 0..1 scale

With Slab preprocessing and matte(): 0.20% mean difference to the same-size PyTorch matte (1.09% on the worst photo). On identical input tensors: 0.094%. Against MODNet's own script, which keeps the aspect ratio: 6.5% (25.5% on the worst photo).

Before you choose this model

  • Made for photos of people; other subjects are not reliably separated.
  • The pack runs at a fixed 512 x 512 and stretches the photo to it. MODNet's own script keeps the aspect ratio, and its mattes differ from the pack's by 6.5% on average across our photos (up to 26% on one). Converting at your camera's aspect ratio removes most of that: at 512 x 672 for 4:3 landscape photos, the difference dropped to 0.45%.
  • Crowds and people far from the camera are separated less cleanly.