Number of Layers

Depth of the U-Net encoder/decoder structure.

This sets how many encoding/decoding stages the network has. Each stage captures patterns at different scales.

Fewer layers (2)

• Smaller receptive field

• Faster training

• Good for local noise patterns

More layers (3-4)

• Larger receptive field

• Better for patterns spanning many pixels

• More memory and compute time

Recommendations

• 2-3 layers for random noise

• 3-4 layers for structured noise with larger patterns

• Consider patch size — more layers need larger patches to be effective