| --- |
| license: apache-2.0 |
| library_name: coreml |
| pipeline_tag: image-to-image |
| tags: [inpainting, lama, coreml, apple-silicon, object-removal] |
| base_model: smartywu/big-lama |
| --- |
| |
| # LaMa (big-lama) for Core ML — object removal |
|
|
| What `img erase` in the arraypress tools downloads: give it a picture and a mask of what to |
| remove, and LaMa fills the hole plausibly. 800×800 image and mask in, image out; outside the |
| mask the input passes through untouched. |
|
|
| ## Provenance and verification |
|
|
| - **Weights:** big-lama from [advimman/lama](https://github.com/advimman/lama) (Apache-2.0), |
| via [smartywu/big-lama](https://huggingface.co/smartywu/big-lama) — the download the official |
| README gives (`curl -LJO https://huggingface.co/smartywu/big-lama/resolve/main/big-lama.zip`). |
| - **This package:** the Core ML conversion published at `jerhoads/lama-coreml` |
| (commit `da9859ec149712a4bba6dfb2946f097b026bc8ec`), which states no licence of its own. |
| Re-hosted unmodified (file hashes identical) so the fleet does not depend on a third-party repo. |
| - **Verified against the official weights** (`verify_against_official.py`): the original PyTorch |
| generator (`FFCResNetGenerator`, `models/best.ckpt`) and this package run on CPU on three |
| scenes (gradient + shapes, noise texture, a stock photo) × two masks: |
| - inside the mask: **58.9 dB PSNR** in every case — 8-bit rounding, i.e. the same model; |
| - outside the mask: identical to the input (max difference 0). |
| - Conversion metadata: coremltools 7.2 from a TorchScript trace (torch 2.2.0), ML Program. |
|
|
| ## Licence |
|
|
| Apache-2.0 — see `LICENSE` and `NOTICE`. |
|
|
