| --- |
| license: mit |
| library_name: onnxruntime |
| pipeline_tag: image-segmentation |
| tags: |
| - background-removal |
| - image-matting |
| - onnx |
| - webgpu |
| - fp16 |
| --- |
| |
| # Lucida Web ONNX (experimental) |
|
|
| This repository contains an experimental ONNX/WebGPU conversion of **Lucida** for personal |
| validation in a web browser. It is not an official release of the original project and is not |
| affiliated with or endorsed by the original author. |
|
|
| ## Original project and acknowledgements |
|
|
| The model architecture, trained weights, and the core background-removal capability come from the |
| original **[egeorcun/lucida](https://github.com/egeorcun/lucida)** project by egeorcun. Lucida is a |
| BiRefNet-based image matting model released under the MIT License. |
|
|
| Many thanks to egeorcun and all upstream Lucida/BiRefNet contributors for releasing their work and |
| making this browser experiment possible. Please refer to the original repository for the model's |
| training details, datasets, benchmarks, limitations, license, and citations. |
|
|
| ## About this repository |
|
|
| This model is produced as part of the personal browser/WebGPU experiments maintained at |
| **[tomosud/lucidamod](https://github.com/tomosud/lucidamod)**. |
|
|
| The experiment includes: |
|
|
| - conversion of the original model to a WebGPU-compatible ONNX graph; |
| - mixed precision (mostly FP16, with GridSample retained in FP32); |
| - restructuring of wide Concat nodes for WebGPU storage-buffer limits; |
| - browser-side inference with ONNX Runtime Web/WebGPU; |
| - experimental foreground-colour decontamination for transparent PNG output. |
|
|
| The file `lucida-web-1024-fp16.onnx` uses a fixed `1 x 3 x 1024 x 1024` input. It is intended for |
| technical evaluation and personal verification. Results, compatibility, memory use, and numerical |
| quality may differ from the original PyTorch implementation. |
|
|
| --- |
|
|
| ## 日本語 |
|
|
| このリポジトリは、LucidaをブラウザのONNX Runtime Web/WebGPUで動かすために変換した、 |
| 個人検証用の非公式な実験モデルです。原典プロジェクトの公式リリースではありません。 |
|
|
| モデル構造、学習済み重み、背景除去能力は、egeorcun氏による原典 |
| **[egeorcun/lucida](https://github.com/egeorcun/lucida)** に由来します。モデルとコードを公開し、 |
| この検証を可能にしてくださったegeorcun氏およびLucida/BiRefNetの各コントリビューターに |
| 深く感謝します。学習内容、データセット、ベンチマーク、制限事項、ライセンス、引用情報は |
| 原典リポジトリを参照してください。 |
|
|
| ONNX変換、WebGPU互換化、FP16混合精度化、およびブラウザ動作検証は、個人検証リポジトリ |
| **[tomosud/lucidamod](https://github.com/tomosud/lucidamod)** で行っています。原版PyTorch実装と |
| 完全に同一の品質・互換性を保証するものではなく、技術検証目的の成果物です。 |
|
|
