RolexAvengers/framers-isnet

ISNet general-use (fp16 ONNX) for in-browser background removal

The ISNet "general use" model from DIS — Highly Accurate Dichotomous Image Segmentation (Qin et al., ECCV 2022), prepared for running in a web browser with ONNX Runtime Web (WebGPU, with a WebAssembly fallback).

File

File	Size	SHA-256
isnet-general-use-fp16.onnx	90.4 MB	029b5d76fc41e8854ae62abe9309230f193e2eeb10d3856bbe1c521df46870c2

• Input input_image: float32 [1, 3, 1024, 1024], RGB, pixel / 255 - 0.5.
• Output output_image: float32 [1, 1, 1024, 1024], matte in roughly 0–1 (min-max normalise it, as DIS's inference does).

Provenance

1. Official weights: isnet-general-use.pth, from https://github.com/xuebinqin/DIS (Apache-2.0).
2. ONNX export: isnet-general-use.onnx from the rembg project's release https://github.com/danielgatis/rembg/releases/tag/v0.0.0 (sha256 60920e99c45464f2ba57bee2ad08c919a52bbf852739e96947fbb4358c0d964a).
3. This file: the 11 deep-supervision side outputs removed (only the final matte is kept), then converted to float16 weights with float32 inputs and outputs (onnxconverter_common.float16, keep_io_types=True). Checked with onnx.checker.

Licence

Apache-2.0, as the original DIS weights — see LICENSE.

Citation

@InProceedings{qin2022,
  author    = {Xuebin Qin and Hang Dai and Xiaobin Hu and Deng-Ping Fan and Ling Shao and Luc Van Gool},
  title     = {Highly Accurate Dichotomous Image Segmentation},
  booktitle = {ECCV},
  year      = {2022}
}