Inpainting (MLX)
Collection
Apple-MLX fp16 inpainting / object-removal models (LaMa Apache-2.0 + MI-GAN MIT). Loaded by mlx-lama-swift. • 5 items • Updated • 1
mlx-community/LaMa-bf16
Superseded — use mlx-community/LaMa-fp32. These bf16 weights cast large fills in colour photos pink/magenta: Big-LaMa can't take bf16 weight rounding (max_abs 7.5e-2 vs PyTorch on a colour fixture; fp32 5.0e-6). mlx-lama-swift ≥ 0.3.1 loads LaMa-fp32.
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.
Model size
51.1M params
Tensor type
BF16
·
Hardware compatibility
Log In to add your hardware
Quantized