gumka/lama-coreml

LaMa for Core ML

Big LaMa, the large-mask inpainting model, as a Core ML package, as used by the Erase tool of Gumka on macOS. It fills a masked area with what fits the picture around it. Its Fourier convolutions see the whole crop at once, so it handles large areas and repeating structure (brickwork, fences, foliage).

Package	Input	Output
LaMa.mlpackage	image 800 × 800 BGRA, 8-bit; mask 800 × 800, one channel, 8-bit (non-zero fills)	output 800 × 800 BGRA, 8-bit

• The size is fixed. Gumka crops the picture around the mask, with context on every side, scales the crop to 800 × 800, and blends the fill back into the full-resolution picture.
• The weights are float32. Gumka runs it on CPU and GPU: about 0.7 s per run on a MacBook Air (Apple Silicon). The outputs of CPU, GPU and Neural Engine agree within 1/255.
• The ONNX version, for Linux, is gumka/lama-onnx.

Source and license

• Model: LaMa, "Resolution-robust Large Mask Inpainting with Fourier Convolutions", Roman Suvorov et al., 2021, https://arxiv.org/abs/2109.07161, https://github.com/advimman/lama (Apache-2.0).
• Core ML conversion: Daisuke Majima (john-rocky), from CoreML-Models, published at mlboydaisuke/LaMa-CoreML (Apache-2.0).
• This repository holds the package of mlboydaisuke/LaMa-CoreML at revision 5ed76e3799ab, unchanged; Gumka's models.toml pins each file's size and SHA-256. License: Apache-2.0, see LICENSE.