birefnet-toonout-onnx

BiRefNet-ToonOut — ONNX

ONNX export of joelseytre/toonout (ToonOut: BiRefNet fine-tuned for anime background removal — Muratori & Seytre, MIT, paper, code), converted so the model runs anywhere ONNX Runtime does — Node.js included, no Python.

Converted with torch.onnx.export (opset 17) plus deform_conv2d_onnx_exporter for the decoder's deformable convolutions. Conversion script: experiments/003-toonout-onnx in the sprited repo.

Files

• birefnet-toonout-fp16.onnx — fp16 weights, fp32 inputs/outputs (no Float16Array juggling in JS). This is what sprited downloads.
• birefnet-toonout.onnx — fp32 original export.

I/O

Preprocessing: plain resize to 1024×1024, RGB / 255, ImageNet normalization (mean 0.485, 0.456, 0.406, std 0.229, 0.224, 0.225). Postprocessing: resize mask back to the source size, use as alpha.

Usage (Node.js)

import ort from "onnxruntime-node";
const session = await ort.InferenceSession.create("birefnet-toonout-fp16.onnx");
const { mask } = await session.run({ image: new ort.Tensor("float32", chw, [1, 3, 1024, 1024]) });

Or just npx sprited gen char — sprited uses this model for matting by default.