sapienkit/LaMa-ONNX

LaMa Inpainting — ONNX (512×512)

LaMa (Resolution-robust Large Mask Inpainting with Fourier Convolutions, WACV 2022) exported to ONNX for object removal / inpainting. Fixed 512×512 input.

I/O contract

• Inputs
  • image — float32[1, 3, 512, 512], RGB, normalised /255
  • mask — float32[1, 1, 512, 512], 1 = region to erase, 0 = keep
• Output
  • output — float32[1, 3, 512, 512], RGB already in [0, 255] (no *255 needed)

File

• lama_fp32.onnx — SHA-256 1faef5301d78db7dda502fe59966957ec4b79dd64e16f03ed96913c7a4eb68d6

Credits & License

Released under the Apache License 2.0.

• Original model: LaMa — advimman/lama (Apache-2.0).
• ONNX export based on Carve/LaMa-ONNX (Apache-2.0).
• Trained on the Places2 dataset (CC-BY 4.0) — attribution to the Places2 authors.

If you use this model, please cite the original LaMa paper:

@inproceedings{suvorov2022resolution,
  title     = {Resolution-robust Large Mask Inpainting with Fourier Convolutions},
  author    = {Suvorov, Roman and others},
  booktitle = {WACV},
  year      = {2022}
}