CoreMLaMa: LaMa for Core ML

This repo contains a script for converting a LaMa (aka cute, fuzzy 🦙) model to Apple's Core ML model format. More specifically, it converts the implementation of LaMa from Lama Cleaner.

This repo also includes a simple example of how to use the Core ML model for prediction. See Sample.

Conversion Instructions

1. Create a Conda environment for CoreMLaMa:

   conda create -n coremlama python=3.10 # works with mamba, too
   conda activate coremlama
   pip install -r requirements.txt

2. Run the conversion script:

   python convert_lama.py

This script will download and convert Big LaMa to a Core ML package named LaMa.mlpackage.

iOS Deployment Problems

The Core ML model this script produces was designed for macOS deployments. It runs well on macOS, on the GPU. I have received several reports of unsuccessful attempts to run this model on iOS, especially with fp16 precision on the Neural Engine. Conversely, I have not received any reports of successful deployments to iOS.

It may very well be possible to run this model on iOS with some tuning in the conversion process. I simply have not attempted this. I would very much welcome a PR and give credit to anyone who is able to convert this model and run it with great results on iOS.

Acknowledgements and Thanks

Thanks to the authors of LaMa:

[Project page] [arXiv] [Supplementary] [BibTeX] [Casual GAN Papers Summary]

CoreMLaMa uses the LaMa model and supporting code from Lama Cleaner. Lama Cleaner makes this project much simpler.