BiRefNet-fp16 (MLX)

mlx-community/BiRefNet-fp16 is an fp16 MLX conversion of ZhengPeng7/BiRefNet (MIT) — a Swin-L + ASPP-Deformable foreground segmentation / matting model at 1024×1024, producing a single-channel soft-alpha matte (white = foreground). The fast, general-purpose tier.

Parity: IoU 0.9905 vs the PyTorch reference (zero unused keys). fp16 runtime validated for production matting quality.

Use (Swift / MLX)

Loaded by mlx-birefnet-swift — the vendored BiRefNet core plus a conformant MLXEngine matting ModelPackage:

import BiRefNet
let pipeline = try BiRefNetPipeline.fromPretrained("model.safetensors", dtype: .float16)  // inputSize 1024
let matte = try pipeline(cgImage).maskCGImage()   // source-resolution soft-alpha

Converted from the official PyTorch checkpoint via the package's birefnet-convert (PyTorch NCHW → MLX NHWC; 754 → 687 tensors). Single-file model.safetensors. See also the higher-res tier mlx-community/BiRefNet_HR-matting-fp16.