imagingforgood/lama-512-onnx

LaMa (Big-LaMa) 512, ONNX for WebGPU

An unchanged copy of an ONNX conversion of Big-LaMa, kept by Imaging for Good so that browser-perception can rely on it. We did not train, convert or modify this model.

Origin

• Model: LaMa (Big-LaMa) by Roman Suvorov et al., Samsung AI Center, Apache-2.0. Code: advimman/lama. Paper: Suvorov et al., Resolution-robust Large Mask Inpainting with Fourier Convolutions, WACV 2022 (arXiv:2109.07161).
• ONNX conversion: lama_512.onnx by Artiva, from artiva-ai/modelos at commit 2f5528556abb9a543bc2af8f7405b0466bdca82b. Per its model card: fixed 512×512 input, and the FFT of the Fourier units written with MatMul/Cos/Sin (Carve-Photos/lama, FourierUnitJIT) so it runs on WebGPU, checked against PyTorch on the same inputs.

File

File	Bytes	SHA-256
lama_512.onnx	207,594,478	59bfc738a9e9aa419d027aae39f9ed9ad8c36331c780756c45cf83c3d54ee2e6

Contract

• image [1, 3, 512, 512] float32, RGB / 255; mask [1, 1, 512, 512] float32, 1 = fill.
• output [1, 3, 512, 512] float32, 0..255. Unmasked pixels aren't guaranteed unchanged; paste back only the masked ones.

Training data

Big-LaMa was trained on Places2, whose terms of use are for non-commercial research and education. The Apache-2.0 licence covers the code and weights; check that the training data's terms fit your use.

Licence

Apache-2.0, as the original model; see LICENSE (the notice from artiva-ai/modelos and LaMa's licence text).