birefnet-lite-onnx
Mirror of onnx/model_fp16.onnx from onnx-community/BiRefNet_lite-ONNX
(MIT; upstream model ZhengPeng7/BiRefNet_lite, code https://github.com/ZhengPeng7/BiRefNet).
Mirrored on 2026-11-20 from upstream commit de15b22ba131738a16dff04aab8bdf8dc32e3ac1. The file is byte-identical to upstream; only the files an app loads are kept here (the fp32 variant is left out).
Files
| File | Bytes | SHA-256 |
|---|---|---|
model_fp16.onnx |
114538221 | d39b897ceb16ae654c1731f3dba0cf9b368d9cae74b5a57459b455cc8bfec402 |
preprocessor_config.json |
391 | (upstream, documents the preprocessing below) |
Tensors (checked with onnxruntime 1.24.1, CPU)
| Name | Shape | dtype | |
|---|---|---|---|
| input | input_image |
[1, 3, 1024, 1024] (NCHW) |
float32 |
| output | output_image |
[1, 1, 1024, 1024] |
float32 |
The weights are stored as fp16, the graph inputs and outputs are float32, so no host-side half conversion is needed.
Pre- and post-processing
- Resize the RGB image to 1024x1024 (bilinear), scale to 0..1, normalise with ImageNet mean
[0.485, 0.456, 0.406]and std[0.229, 0.224, 0.225]. - The output is logits (measured range about -24 .. +20 on a synthetic test image): apply a sigmoid, then resize the 1024x1024 map back to the source size. The result is the alpha matte.
- Measured CPU latency on a desktop (onnxruntime-python, one 1024x1024 run): about 20 s on a loaded machine; budget for phones is in the ProductStage spec.
Provenance
No conversion was run here: the file was downloaded from the upstream repository with curl and re-uploaded unchanged.
Citation: Zheng et al., "Bilateral Reference for High-Resolution Dichotomous Image Segmentation", CAAI AIR 2024.
Model tree for ipsilondev/birefnet-lite-onnx
Base model
ZhengPeng7/BiRefNet_lite