Big-LaMa 512 for the LiteRT GPU delegate

Big-LaMa (Suvorov et al., "Resolution-robust Large Mask Inpainting with Fourier Convolutions", WACV 2022) as a 512 x 512 LiteRT (TensorFlow Lite) model where every operation runs on the Android GPU delegate.

  • Inputs: RGB image 1x512x512x3 in 0..1 and erase mask 1x512x512x1 (exactly 0 or 1; 1 = fill).
  • Output: RGB 1x512x512x3 in 0..255.
  • Weights are FP32; run with precision loss allowed (FP16) for speed.

Converted from the ONNX export at https://huggingface.co/AUXOUT-TEAM/agit-lama-inpainting (revision 5646fb0bbd8887fac005e0bd3be7047f2211f74b), itself derived from https://huggingface.co/smartywu/big-lama. The Fourier units' DFT matrices are folded for the fixed size, their real/imaginary stacking is rewritten in 4D, and the DFT matrix products are expressed as 1x1 convolutions. Output matches the source ONNX within 1e-4 on the CPU.

License

Apache-2.0, as the original code and weights: https://github.com/advimman/lama

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support