fp16 MLX conversion of ZhengPeng7/BiRefNet_HR-matting (Swin-L, 2048) — soft-alpha matting

144d317 verified 21 days ago

1.59 kB

license: mit
base_model: ZhengPeng7/BiRefNet_HR-matting
library_name: mlx
tags:
  - mlx
  - image-segmentation
  - matting
  - background-removal
  - birefnet
pipeline_tag: image-segmentation

BiRefNet_HR-matting-fp16 (MLX)

mlx-community/BiRefNet_HR-matting-fp16 is an fp16 MLX conversion of ZhengPeng7/BiRefNet_HR-matting (MIT) — the same Swin-L + ASPP-Deformable architecture run at 2048×2048 for the crispest dense-hair detail. The high-resolution "best" matting tier (best all-rounder: crispest fine hair while retaining thin structures like whiskers).

Parity: IoU 0.9905 vs the PyTorch reference (loads through the identical converter + model as the general weights, zero code change). fp16 runtime validated for production matting quality. ~2 s/image at 2048 on Apple Silicon (≈18 GB peak — a pro-tier footprint).

Use (Swift / MLX)

Loaded by mlx-birefnet-swift:

import BiRefNet
var cfg = BiRefNetConfig.swinLargeDefault; cfg.inputSize = (2048, 2048)
let pipeline = try BiRefNetPipeline.fromPretrained("model.safetensors", dtype: .float16, config: cfg)
let matte = try pipeline(cgImage).maskCGImage()   // source-resolution soft-alpha

Converted from the official PyTorch checkpoint via the package's birefnet-convert. Single-file model.safetensors. The fast tier is mlx-community/BiRefNet-fp16.