IS-Net general-use (ONNX) — ClearCut mirror
A self-controlled mirror of the IS-Net general-use ONNX model, hosted to provide a stable source
for the High-Quality background-removal path in
ClearCut — a zero-backend, fully client-side AI background
remover that runs this model in the browser via onnxruntime-web (WASM).
The fp32 weights are redistributed unchanged under their original Apache-2.0 license. The -q8
file is a uint8 dynamic-quantization of those same weights, produced for faster, lighter in-browser
CPU (WASM) inference.
Files
| File | Size (bytes) | SHA-256 | Notes |
|---|---|---|---|
isnet-general-use.onnx |
176213804 | 4c56bbc21588459dda11efba5a4a8ee163969da109ae170fb1988c1c2ea4a90a |
Verbatim Apache-2.0 fp32 weights (reference). |
isnet-general-use-q8.onnx |
44436071 | feed6f32a5e707ca7e939576b2d891b23fb9eb4114749657a5efc64e8651e43a |
uint8 dynamic-quantized (onnxruntime.quantization.quantize_dynamic, QUInt8). Served by ClearCut HQ — ~4× smaller, ~1.8× faster on the WASM CPU EP, mask quality near-identical to fp32 (IoU ≈ 0.94 on a fine-edge test, fine detail preserved). |
Input input (1×3×1024×1024, normalized (pixel − 128) / 256); output output (saliency 0..1).
Both files keep fp32 input/output tensors, so they are interchangeable at the application boundary.
Attribution
- Model / paper: X. Qin et al., "Highly Accurate Dichotomous Image Segmentation", ECCV 2022.
- Original pretrained weights: xuebinqin/DIS (official repo).
- ONNX export mirrored from: x-Liola-x/isnet-general-use-onnx.
License
Apache-2.0 — same as the upstream model and ONNX export. The quantized -q8 file is a derived
redistribution of the same Apache-2.0 weights. See the original repositories above for the full
license text and citation.