LaMa (ONNX) — texture inpainting for QtMeshEditor
Re-hosted ONNX export of LaMa (large-mask inpainting), used by QtMeshEditor to fill masked regions of a texture with plausible, seamlessly continued content (issue #1017).
- Paper: Suvorov et al., "Resolution-robust Large Mask Inpainting with Fourier Convolutions", WACV 2022.
- Licence: Apache-2.0 (code and weights).
- Source graph:
Carve/LaMa-ONNX.
Which file this is
The upstream repo publishes two graphs. The plainly-named lama.onnx there is
broken — it fails ONNX Runtime shape inference on a DFT node (LaMa's
Fourier convolution) and cannot be loaded at all:
Op (DFT) [ShapeInferenceError] one-sided DFT requires real input
This repo hosts the working lama_fp32.onnx under the plain name
lama.onnx, which is what the application downloads.
I/O contract
Measured, not documented upstream:
| name | dtype | shape | range | |
|---|---|---|---|---|
| input | image |
float32 | [1,3,512,512] |
RGB scaled to [0,1] |
| input | mask |
float32 | [1,1,512,512] |
1 = inpaint, 0 = keep |
| output | output |
float32 | [1,3,512,512] |
RGB already in 0..255 |
Note the asymmetry — in [0,1], out 0..255. Feeding an un-scaled image
produces a near-white result (measured mean 246.6 vs a correct 128.6) with no
error raised.
Spatial dims are fixed, so larger textures are processed in overlapping 512² tiles with a feathered seam blend.
Measured behaviour
On a thin-seam mask (the UV-bleed repair case), against ground truth the model never saw:
- reconstruction MAE 9.15 / 255 (3.6%) inside the mask
- leakage MAE 0.0000 outside the mask — unmasked pixels are preserved exactly
Usage
Downloaded automatically on first use by QtMeshEditor. Also available in the
aggregate repo
QtMeshEditor-models
under pbr/lama.onnx, which is the path the runtime fetches.
qtmesh material --texture albedo.png --inpaint --mask holes.png -o fixed.png