LaMa (ONNX) — object-removal inpainting

lama_fp32.onnx — an ONNX export of LaMa (big-lama) for image inpainting (erasing a masked object and filling the hole). Runs client-side with onnxruntime-web on the WASM or WebGPU execution providers.

A verbatim mirror of Carve/LaMa-ONNX's lama_fp32.onnx, hosted so a single, immutable, CORS-enabled copy backs the Edge Tools "Remove Object" tool.

Inputs / outputs

• Inputs: image float32 [1, 3, 512, 512], mask float32 [1, 1, 512, 512]
• Output: output float32 [1, 3, 512, 512] — the inpainted image
• RGB, channels-first (CHW). Image values in 0..1; mask polarity: 1 = erase (the hole), 0 = keep. The output is the raw fill (0..255), composited over the original by the caller so only the hole changes. Spatial size is fixed at 512×512 (resize to it, then back).

Provenance & license

• Algorithm / weights: LaMa — Resolution-robust Large Mask Inpainting with Fourier Convolutions (Suvorov et al., WACV 2022), advimman/lama. The big-lama weights are released under the Apache-2.0 license.
• ONNX export: Carve/LaMa-ONNX (lama_fp32.onnx).
• The full Apache-2.0 LICENSE is included in this repo.