ISNet (anime-seg) for onnxruntime-web / WebGPU

A conversion of skytnt/anime-seg (SkyTNT Anime Segmentation, ISNet, Apache-2.0) so it runs in the browser on WebGPU. All credit for the model goes to its author.

What changed from the original weights:

  • Exported from the original model.safetensors to ONNX, opset 17, fixed input 1x3x1024x1024.
  • Every MaxPool has ceil_mode switched off. With a 1024x1024 input every pooling stage sees an even size, so the result is identical, and onnxruntime-web (WebGPU) does not support ceil_mode.
  • Output is the raw logit map (logits, 1x1x1024x1024). Apply a sigmoid yourself. The input is named input.

Checked against the original PyTorch model on a real picture: IoU 1.0, 0 differing pixels (onnxruntime CPU, and onnxruntime-web on both WebGPU and WASM).

Pre-processing: RGB in [0, 1], resized to fit 1024 and zero-padded to a square, NCHW float32.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sonwnguyen13/vidw-isnet-web

Base model

skytnt/anime-seg
Quantized
(2)
this model