isnet-general-use (ONNX, fp16 / q8)

• Source ONNX export: the isnet-general-use.onnx distributed by the rembg project (MIT); original weights and architecture by Xuebin Qin et al. (DIS, Apache-2.0).
• Modifications: the 11 auxiliary graph outputs (side stages d2-d6 and intermediate feature taps) were removed, keeping only output_image (stage-1 sigmoid probability map). fp16 and dynamic-int8 (q8) variants were generated with the official Transformers.js conversion scripts.

I/O

• Input input_image: float32 [1, 3, 1024, 1024], RGB, normalized as x / 255 - 0.5 (mean 0.5, std 1.0).
• Output output_image: float32 [1, 1, 1024, 1024], sigmoid probabilities. Min-max normalize per image for an alpha matte (as done by rembg).

Files

• onnx/model.onnx - fp32 (aux outputs stripped)
• onnx/model_fp16.onnx - fp16
• onnx/model_quantized.onnx - dynamic int8 (q8)