mlx-community/LaMa-bf16

Big-LaMa FFC inpainting (quality tier), converted to Apple MLX (bf16) for the mlx-lama-swift Swift package (MLXEngine imageInpaint / Forge Erase). From-scratch MLX-Swift port of advimman/lama; parity-locked vs PyTorch (predicted max_abs 3.2e-5 fp32).

bf16, not fp16: LaMa's Fast Fourier Convolution produces high-magnitude bottleneck activations (~1e3); fp16 weight rounding collapses the model (mean err 0.55), while bf16's wider exponent range preserves it (mean 4e-3, visually identical to fp32). Use bf16 or fp32 — not fp16.

import LaMa
let inpainter = try LaMaInpainter.fromPretrained(weightsPath, dtype: .bfloat16)
let filled = inpainter(sourceCGImage, mask: maskCGImage)  // white mask = remove

Weights: Apache-2.0 (advimman/lama). Port: MIT.