NeuraTable is a framework and tool to augment open source photography workflows with machine learning tools.
NeuraTable is still in a very early stage of development. It's more of a proof of concept for now, but I hope to grow this into a useful and easy to use application in the future.
NeuraTable can run neural networks in the ONNX format. It uses the Wonnx runtime to run these models on all platforms where Vulkan is available.
As of right now, NeuraTable does not ship any ONNX models. The inference logic is mostly inspired from nind-denoise, and my plan is to ship a pretrained nind-denoise model for version 1.
If you want to give the current state of the project a go, you can try to convert one of the nind-denoise UNet snapshots to
ONNX and run that through the neuratable_run_onnx binary.
NeuraTable currently provides an executable called neuratable_run_onnx.
To build this executable, you'll need a stable rust toolchain and Vulkan available on your system.
If you use NixOS, you can just use nix develop to get a shell with the dependencies in place.
To run a network on an example, use cargo run --release -- <PATH_TO_MODEL.onnx> <PATH_TO_INPUT.jpg> <PATH_TO_OUTPUT.jpg>