# Neural LOD Selector
Training a small neural network using C++, Metal and Pytorch to predict the pixel error of different LOD levels. The network is embedded in a compute shader (HLSL) to run neural LOD selection and indirect drawing on the GPU.
The model automatically learns the relationships between different input features (e.g. object distance, screen size, velocity, FOV...) and the loss in image quality for each LOD level. We can then use the predicted pixel error to pick an optimal LOD.
Training the network is done using a standalone C++ application:
Training LOD levels: pixel error highlighted in green
We can use the model prediction to apply content-aware LODs based on the mesh characteristics. In this example, the delicate Sword asset maintains higher quality LODs at a distance, while the Rock uses more aggressive LODs. Feature selection is used to train the model based on combination of different parameters. By including the camera FOV as an input feature, the model can also switch to higher quality LODs when zooming on objects:
To learn more, check the Github repo: https://github.com/eldnach/neural-lod