| --- |
| license: apache-2.0 |
| library_name: onnx |
| pipeline_tag: image-segmentation |
| tags: |
| - background-removal |
| - onnx |
| - onnxruntime-web |
| - dichotomous-image-segmentation |
| --- |
| |
| # ISNet general-use (fp16 ONNX) for in-browser background removal |
|
|
| A half-precision (fp16) copy of the **ISNet general-use** model, used by techyCoderWorld's Image |
| Studio to remove image backgrounds entirely in the visitor's browser with ONNX Runtime Web |
| (WebGPU or WebAssembly). Images never leave the device. |
|
|
| ## Source and licence |
|
|
| - Model: **ISNet** from *Highly Accurate Dichotomous Image Segmentation* (DIS), Xuebin Qin, |
| Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, Luc Van Gool (ECCV 2022). |
| Code and weights: https://github.com/xuebinqin/DIS, released under the **Apache License 2.0**. |
| - ONNX export used as the starting point: `isnet-general-use.onnx` distributed by the rembg |
| project (https://github.com/danielgatis/rembg), SHA-256 `60920e99c45464f2ba57bee2ad08c919a52bbf852739e96947fbb4358c0d964a`. |
|
|
| ## Changes made |
|
|
| - Weights and internal computation converted from float32 to float16 with |
| `onnxconverter-common` 1.14.0 (`keep_io_types=True`), so inputs and outputs stay float32. |
| - No retraining or other changes. Output matches the float32 model to within fp16 precision. |
|
|
| ## File |
|
|
| | File | SHA-256 | Size | |
| |---|---|---| |
| | `model_fp16.onnx` | `d27e782d3bec1c787c5b6c612a1aae8526373a721540d4046939e8a75d2db949` | 90662600 bytes | |
|
|
| ## Input and output |
|
|
| - Input `input_image`: float32 `[1, 3, 1024, 1024]`, RGB, each pixel divided by the image's |
| maximum value, then minus 0.5. |
| - Output: float32 `[1, 1, 1024, 1024]`; min-max normalise to get the foreground mask (0..1). |
|
|
| Licensed under the Apache License, Version 2.0: https://www.apache.org/licenses/LICENSE-2.0 |
|
|
