Number of Features

Base number of convolutional filters in the network.

This controls the initial number of feature maps (filters) in the first layer of the U-Net. More features mean a more powerful model.

Fewer features (32-48)

• Smaller, faster model

• Less memory usage

• May miss complex noise patterns

More features (64-128)

• More powerful model

• Better at learning complex patterns

• Requires more GPU memory and training time

Recommendations

• 64 is a good default

• Reduce to 32 if memory is limited

• Increase to 96-128 for complex structured noise