BiRefNet lite — 512px, WebGPU-friendly ONNX

Background removal model used in the Dayfolie carousel editor (runs fully in the browser with onnxruntime-web 1.30, WebGPU or WASM).

Source weights: ZhengPeng7/BiRefNet_lite (MIT)
Re-exported at 512×512 input, fp16 weights with float32 input/output.
Large Concat/Split ops were split into smaller ones and ReduceSum replaced with Add so the graph runs on WebGPU.
Input: input_image float32 [1,3,512,512] (ImageNet mean/std). Output: mask logits [1,1,512,512].

Original work: Zheng Peng et al., Bilateral Reference for High-Resolution Dichotomous Image Segmentation (CAAI AIR 2024). MIT License, © ZhengPeng7.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for dayfolie/birefnet-lite-512-webgpu

Base model

ZhengPeng7/BiRefNet_lite

Quantized

(29)

this model