tiper-penta/isnet-general-use-fp16

ISNet general-use — fp16 ONNX

Background-removal (salient object segmentation) model for in-browser inference with ONNX Runtime Web.

Source

• Model: IS-Net from "Highly Accurate Dichotomous Image Segmentation" (Qin et al., ECCV 2022) — https://github.com/xuebinqin/DIS (Apache-2.0)
• ONNX export: isnet-general-use.onnx from rembg — https://github.com/danielgatis/rembg (MIT)
  • Download: https://github.com/danielgatis/rembg/releases/download/v0.0.0/isnet-general-use.onnx
  • SHA-256: 60920e99c45464f2ba57bee2ad08c919a52bbf852739e96947fbb4358c0d964a

Modifications

1. Removed the auxiliary side outputs; only the final output_image output is kept.
2. Converted weights to float16 with onnxconverter-common (keep_io_types=True, so input/output remain float32).

No retraining. Output matches the fp32 original (mean alpha difference < 0.02% on test images).

I/O

• Input input_image float32 [1, 3, 1024, 1024], RGB, pixel / max(pixel) - 0.5
• Output output_image float32 [1, 1, 1024, 1024] foreground probability (min-max normalize before use)

License

Apache-2.0 (see LICENSE), following the original DIS / IS-Net release.