Patch Size

Size of the square image patches used for training.

During training, small square patches are extracted from your images. This dropdown lives under Advanced Options, and its default and choices depend on the branch.

The control

• N2V branch: default 64, options 32, 48, 64, 96, 128

• StructN2V branch: default 128, options 32, 48, 64, 96, 128, 256

The StructN2V branch defaults larger so the structural mask has enough spatial context around each pixel.

Smaller patches

• Train faster with less memory

• Good for fine-grained noise

Larger patches

• Capture more context

• Better for structured noise with larger periods

• Require more GPU memory

Validation

The patch size is checked live against your image and the network geometry:

• It cannot exceed the smaller image dimension. Options larger than your image are disabled automatically.

• The extract size (patch size + 32 pixels of overlap-tile padding) must be divisible by 4. An invalid value is flagged and the Next button is blocked until it is corrected.

Recommendations

• Keep the branch default (64 for N2V, 128 for StructN2V)

• Reduce it if you hit memory errors or your images are small