For_my_PhotoCraft
Models stored in this repository are used by PhotoCraft, a native Rust image editor, and PhotoCraft downloads them automatically from here when they are missing on the user's machine.
Files
| File | Size | SHA-256 | Used by |
|---|---|---|---|
big-lama.onnx |
208,044,816 bytes (198.4 MiB) | 1faef5301d78db7dda502fe59966957ec4b79dd64e16f03ed96913c7a4eb68d6 |
PhotoCraft βΊ LaMa Cleaner (paint.lamaCleaner) |
big-lama.onnx is an unmodified, byte-identical copy of lama_fp32.onnx from Carve/LaMa-ONNX. Identity is proven by the SHA-256 above, which equals the LFS hash of the source file on the upstream repository. The model file itself was not changed in any way.
The model: big-lama (LaMa inpainting)
- Fixed 512Γ512 input:
image [1,3,512,512]float in 0..1 (RGB, CHW) andmask [1,1,512,512]binary (0 or 1); the output is planar RGB in 0..255. - Opset 17, exported with
torch.onnx.export, with a customFourierUnitJITimplementation that produces results identical to the original PyTorch model (per the upstream port notes). The input resolution is fixed at 512Γ512; PhotoCraft reaches arbitrary resolutions with a crop-and-pad strategy around that fixed core. - The inference runs in a helper process through ONNX Runtime with a CUDA β DirectML β CPU execution-provider chain.
Licenses and attribution
Everything in this repository is distributed under the Apache License 2.0. A copy of the license text is included in LICENSE (taken verbatim from the original LaMa repository).
- Original model: LaMa β advimman/lama, licensed under Apache-2.0, Β© 2021 Samsung Research (Samsung AI Center). Paper: Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Nikita Gontcharovski, Konstantin Sofiiuk, Victor Lempitsky, "Resolution-robust Large Mask Inpainting with Fourier Convolutions", WACV 2023.
- ONNX port:
lama_fp32.onnxβ Carve/LaMa-ONNX, Apache-2.0. This is the file redistributed here, byte-identical and unmodified. - Redistribution statement (Apache-2.0 Β§4): the license text is provided in LICENSE; the copyright, patent, trademark and attribution notices of the original work are retained in this card; the redistributed file carries no modifications, so no change notices are required.
Disclaimer
This repository and its maintainer are not affiliated with, and not endorsed by, Samsung, Sber, or Carve. The models are redistributed as-is, under their Apache-2.0 licenses, without any warranty.