LaMa Inpainting Model

HG Space: https://huggingface.co/spaces/Carve/LaMa-Demo-ONNX

Description

1. lama_fp32.onnx (RECOMMENDED)

1. Custom FourierUnitJIT: A custom FourierUnitJIT implementation is used since the original cannot be directly ported to ONNX without overhead. The result is identical to the original model.
2. Fixed Input Shape: The input shape is fixed at 512x512 pixels. Although dynamic input shapes are possible, they would require resolving issues with dynamic padding in the irfft and rfftn functions in ffc.py.
3. Opset Version 17: This model uses opset version 17.
4. Exportable to TensorRT: The model can be successfully used in TensorRT, etc.

if you need other resolution - export it using our jupyter notebook

2. lama.onnx (NOT RECOMMENDED)

1. Custom DFT irfftn Logic: Uses a custom irfftn ONNX logic (patched onnxscript).
2. Fixed Input Shape: The input shape is fixed at 512x512 pixels.
3. Opset Version 18: This model uses opset version 18.
4. Performance: The model works slowly due to issues with torch.onnx.dynamo_export and optimization of the ONNX model.

Resources

• Original repository: advimman/lama
• Repository with custom implementation of exportable LaMa: Carve-Photos/lama

Example

Original image:

lama_fp32.onnx - output:

lama.onnx - output:

Original model output: