smile-forge/big-lama-onnx

big-lama for ONNX Runtime

LaMa's big-lama — Resolution-robust Large Mask Inpainting with Fourier Convolutions — as one ONNX graph for ONNX Runtime on CPU or CUDA: a picture and a mask in, the picture back with what the mask covers filled from what is around it. Scratches, dust, wires, a person or a sign rubbed out of a photo. Built for Nexaflow, whose AI host runs it from C# for its image editor's eraser; usable by anything that runs ONNX Runtime 1.30 or later.

Folder	What	From	Size
inpainter/	big-lama, fp32	advimman/lama at 786f593, weights from smartywu/big-lama as LaMa's README names	209 MB

It takes rgb [1, 3, H, W] from 0 to 1 and mask [1, 1, H, W], 1 where the picture is filled and 0 where it is kept, H and W multiples of 8 from 64 to 4096, and gives inpainted [1, 3, H, W] from 0 to 1: the picture, with the generator's prediction inside the mask — LaMa's predict around its generator, out_key: inpainted.

How it was built

upstream.py fetches LaMa at a pinned commit and keeps only the generator's modules; Build-Lama.ps1 downloads the weights and checks them against their hash, and reads the checkpoint with an unpickler that keeps tensors alone, so nothing it names is imported or run; build_lama.py traces the generator with torch.onnx.export and holds the graph to the original in fp32, on CPU and on CUDA.

The graph differs from the original in one place: its Fourier units take their FFTs as matrix products (fourier_matmul.py). Exported as it is, each unit becomes ONNX DFT operators, which ONNX Runtime's CUDA provider does not implement and its CPU provider loops over in O(n²) for any length that is not a power of two — over two minutes for a 1800×1200 picture. Written as products with cosine and sine matrices built from the input's size, the same sums at the same ortho scale run as any other layer on every provider. The generator with products agrees with the generator with torch.fft to 4×10⁻⁶ — a thousandth of an 8-bit level — and the graph with the original to 8×10⁻⁵, on CPU and CUDA.

lama_reference.py runs the makers' own predict — the dataset's reading, symmetric padding to a multiple of 8, the mask taken where it is over 0, the result clipped and truncated to 8 bits — on SAM 3's sample pictures with a drawn mask each, which is what a port is held to. Nexaflow's C# eraser, through this graph, fills the same pixels as that predict to 58–62 dB PSNR on a GPU and 85 dB on a CPU, and leaves every other pixel as it was.

Using it

As LaMa's predict: read the picture as RGB over 255 and the mask as grey over 255, taken as 1 where it is over 0; pad both at the right and bottom to a multiple of 8 by symmetric mirroring; run the graph; cut the padding off; make each value 8 bits by clipping its product with 255 and truncating.

The generator was trained on 256-pixel crops and holds up to about two thousand pixels a side. Nexaflow works round each part of a mask with as much picture again on every side, at least 512 pixels each way, and brings a region larger than 1536 pixels square down to that before the graph — by an averaging resize — and its fill back up after, by Lanczos. At 1536 square the graph peaks at 4 GB on a GPU and takes 0.1 s on an RTX PRO 6000, 4.6 s on 16 CPU threads.

Licence

big-lama's code and weights are Apache-2.0, a copy of which is LICENSE; so is the recipe here.

Suvorov, Logacheva, Mashikhin, Remizova, Ashukha, Silvestrov, Kong, Goka, Park, Lempitsky. Resolution-robust Large Mask Inpainting with Fourier Convolutions. WACV 2022.