faisal-shohag/fashn-human-parser-onnx

FASHN Human Parser (ONNX, transformers.js)

ONNX export of fashn-ai/fashn-human-parser (SegFormer-B4, 18 classes, 384×576 input) for in-browser use with Transformers.js.

File	Precision	Size	Pixel agreement vs. PyTorch*
onnx/model.onnx	fp32	257 MB	99.86%
onnx/model_fp16.onnx	fp16 (fp32 I/O)	130 MB	99.86%
onnx/model_quantized.onnx	int8 dynamic	67 MB	99.78%

* Argmax label map compared with FashnHumanParser.predict on the example images from the FASHN VTON repo.

Exported with optimum (opset 17); fp16 via onnxconverter-common; int8 via onnxruntime.quantization.quantize_dynamic.

Usage

import { pipeline } from "@huggingface/transformers";

const parser = await pipeline("image-segmentation", "faisal-shohag/fashn-human-parser-onnx", { dtype: "q8" });
const segments = await parser("person.jpg"); // [{ label: "top", mask: RawImage }, ...]

Labels: background, face, hair, top, dress, skirt, pants, belt, bag, hat, scarf, glasses, arms, hands, legs, feet, torso, jewelry.

License

Inherits the NVIDIA Source Code License for SegFormer from the original model: non-commercial use only. All credit for the model goes to FASHN AI.