gumka/lama-onnx

LaMa for ONNX Runtime

Big LaMa, the large-mask inpainting model, as an ONNX graph, as used by the Erase tool of Gumka on Linux. It fills a masked area with what fits the picture around it. Its Fourier convolutions see the whole crop at once, so it handles large areas and repeating structure (brickwork, fences, foliage).

File	Input	Output
lama_fp32.onnx (208 MB)	image float32 [1, 3, 512, 512], RGB from 0 to 1; mask float32 [1, 1, 512, 512], 1 fills	output float32 [1, 3, 512, 512], RGB from 0 to 255

• The size is fixed (only the batch axis is open). Gumka crops the picture around the mask, with context on every side, scales the crop to 512 × 512, and blends the fill back into the full-resolution picture.
• The output is already joined to the input outside the mask, and what is under the mask on input is ignored.
• Opset 17. Gumka runs it on ONNX Runtime's CPU provider: about 1.6 s per erase on a Ryzen 7 8845HS.
• The Core ML version, for macOS, is gumka/lama-coreml.

Source and license

• Model: LaMa, "Resolution-robust Large Mask Inpainting with Fourier Convolutions", Roman Suvorov et al., 2021, https://arxiv.org/abs/2109.07161, https://github.com/advimman/lama (Apache-2.0).
• ONNX export: Carve, published at Carve/LaMa-ONNX (Apache-2.0). It replaces the Fourier unit with one that exports (Carve-Photos/lama); Carve reports the same results as the original.
• This repository holds lama_fp32.onnx of Carve/LaMa-ONNX at revision c3c0c9e46893, unchanged; Gumka's models.toml pins its size and SHA-256. License: Apache-2.0, see LICENSE.