executorch-community/MODNet-ExecuTorch-WebGPU

MODNet portrait matting for ExecuTorch WebGPU

This repository contains an ExecuTorch conversion of https://github.com/ZHKKKe/MODNet for the ExecuTorch WebGPU backend, so the model runs entirely in a web browser on the visitor's GPU. It powers the corresponding demo in the ExecuTorch WebGPU Playground (launching soon on executorch.ai).

• Base model: https://github.com/ZHKKKe/MODNet
• Format: ExecuTorch .pte programs delegated to the WebGPU backend
• Quantization: fp32 (as exported)
• Total size: 24.9 MB

About the original model

A real-time portrait matting network that separates people from the background without a trimap.

• Developed by: Zhanghan Ke et al.
• Paper: MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition
• 8.7M parameters in the released checkpoint
• Photographic portrait matting checkpoint

For training data, evaluation results and full details, see the original model card: https://github.com/ZHKKKe/MODNet.

Intended use and limitations

Cutting people out of photos or camera frames, for example to blur or replace the background.

Trained for portraits: results degrade on non-human subjects, several overlapping people, or unusual poses and lighting.

What changed in this conversion

• Fixed 256x256 input; the matte is upsampled to the frame size outside the program

Inputs and outputs

• Input [1, 3, 256, 256] float32 RGB normalized to [-1, 1]: (x / 255 - 0.5) / 0.5
• Output [1, 1, 256, 256] float32 alpha matte in [0, 1]

Files

File	Size	SHA-256
modnet_webgpu.pte	24.9 MB	0c2ff92e26087e98…

Full checksums are in manifest.json. The playground verifies them after download.

Running it

These programs were exported with the ExecuTorch WebGPU backend (backends/webgpu) and validated with the WebAssembly runtime build used by the ExecuTorch WebGPU Playground (launching soon on executorch.ai). That runtime build is not published yet; until it is, the playground is the supported way to run them. It downloads these files once, caches them in the browser, and runs everything locally.

Requirements: a browser with WebGPU enabled.

Provenance

• Weights: boopathiraj/MODNet @ a967175f8fd63de67a1c161d8daac966d2b57e3a, file modnet/pretrained/modnet_photographic_portrait_matting.ckpt, sha256 7c22235f0925deba15d4d63e53afcb654c47055bbcd98f56e393ab2584007ed8 (from a community mirror of the official release)
• Checked against the official checkpoint: modnet_photographic_portrait_matting.ckpt is byte-identical to the official checkpoint linked from github.com/ZHKKKe/MODNet (sha256 7c22235f0925deba15d4d63e53afcb654c47055bbcd98f56e393ab2584007ed8), and the model code used for the export is identical to the official src/models/ (one import path differs), checked by the project owner on 2026-10-16
• Exported with: executorch 1.5.1+cpu, torch 2.14.1+cpu, torchao 0.18.0+cpu, transformers 5.19.0 (executorch git 3b6068392324)
• Reproduce: python recipes/modnet.py --out out/modnet (recipe in export/)
• Re-running the recipe produces byte-identical files.

Citation

If you use this model, please cite the original work:

@inproceedings{ke2022modnet,
  title={MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition},
  author={Ke, Zhanghan and Sun, Jiayu and Li, Kaican and Yan, Qiong and Lau, Rynson W.H.},
  booktitle={AAAI},
  year={2022}
}

License

This is a derivative of the upstream model and is distributed under the same terms (apache-2.0); see LICENSE. Modifications (conversion to the ExecuTorch format) by Meta Platforms, Inc.