| --- |
| license: apache-2.0 |
| library_name: transformers.js |
| pipeline_tag: image-segmentation |
| tags: |
| - background-removal |
| - onnx |
| - onnxruntime-web |
| - webgpu |
| base_model: onnx-community/ormbg-ONNX |
| --- |
| |
| # ORMBG — compact web build (44 MB) |
|
|
| A compact build of **ORMBG** (Open Remove Background Model, by schirrmacher) for fast in-browser background removal |
| with [Transformers.js](https://huggingface.co/docs/transformers.js) / onnxruntime-web. |
|
|
| Used by the free [Doetra Background Remover](https://doetra.com). |
|
|
| ## What changed vs. `onnx-community/ormbg-ONNX` |
|
|
| - Graph upgraded from opset 11 to opset 17. |
| - **int8 per-channel weights** (`DequantizeLinear`) for convolutions, fp16 for the remaining large tensors, **fp32 |
| compute** — 44 MB instead of 176 MB. Mask difference vs. fp32: mean 0.2/255, 99.93%+ identical after thresholding. |
| - Runs on WebGPU without `shader-f16` support. |
|
|
| ## Usage |
|
|
| ```js |
| import { AutoModel, AutoProcessor } from '@huggingface/transformers' |
| const model = await AutoModel.from_pretrained('Saimon8420/ormbg-web', { dtype: 'fp32', device: 'webgpu' }) |
| const processor = await AutoProcessor.from_pretrained('Saimon8420/ormbg-web') |
| ``` |
|
|
| Input `pixel_values` `[1, 3, 1024, 1024]` in `[0, 1]`; output `alphas` `[1, 1, 1024, 1024]` in `[0, 1]`. |
|
|
| ## Credits & license |
|
|
| - **ORMBG** by Maximilian Schirrmacher — https://github.com/schirrmacher/ormbg (Apache-2.0); ONNX by |
| [onnx-community](https://huggingface.co/onnx-community/ormbg-ONNX). |
| - Redistributed under the same **Apache-2.0** license. |
|
|
