Instance Normalization (InstanceNorm) normalizes each individual example, each individual channel, across only its spatial dimensions — a narrower slice than either BatchNorm or LayerNorm, developed specifically for image style transfer.
Formula
For a single example \(n\) and single channel \(c\), statistics are computed across only that channel's spatial locations (height \(H\) and width \(W\)) — unlike BatchNorm (which pools across the batch, per channel) or LayerNorm (which pools across every channel, per example), InstanceNorm keeps both the example and the channel fixed, normalizing only within that one specific feature map.
Comparing All Three So Far
| BatchNorm | LayerNorm | InstanceNorm | |
|---|---|---|---|
| Fixed dimension(s) | Channel | Example | Example AND channel |
| Pooled dimension(s) | Batch (+ spatial, for images) | All channels (+ spatial) | Spatial only |
Why This Suits Style Transfer
In neural style transfer, a single image's overall contrast and color statistics per channel often encode much of its visual "style." InstanceNorm's per-example, per-channel normalization strips out exactly this instance-specific style information from each feature map — leaving behind content-related structure while discarding the specific contrast/brightness statistics that differ from one instance to the next. This turns out to be a valuable inductive bias when a network needs to reason about content and style as separable concerns, and is exactly why InstanceNorm became closely associated with generative image style-transfer architectures.
Code
import torch
import torch.nn as nn
instance_norm = nn.InstanceNorm2d(num_features=64) # for a (N, C, H, W) image tensor with 64 channels
x = torch.randn(8, 64, 32, 32) # batch of 8, 64 channels, 32x32 spatial
output = instance_norm(x)
print(output.shape) # torch.Size([8, 64, 32, 32]) -- shape unchanged, only statistics normalizedCommon Mistakes
- Applying InstanceNorm to standard classification CNNs by default — it's specifically well-suited to generative/style-transfer contexts where discarding per-instance contrast information is desirable; for typical classification tasks, BatchNorm or GroupNorm are more common choices.
- Confusing InstanceNorm with LayerNorm — both normalize per-example, but InstanceNorm additionally keeps channels separate (normalizing only within each channel's own spatial extent), while LayerNorm pools across all channels together.
Interview Relevance
Q: "Why is Instance Normalization particularly well-suited to style transfer, compared to BatchNorm?" InstanceNorm normalizes each example's each channel independently across only its spatial dimensions, which removes per-instance contrast and brightness statistics — information closely tied to an image's visual "style." This lets a style-transfer network separate style-related statistics from content-related structure more cleanly than BatchNorm, which pools statistics across the batch and would blend information across different images.
Practice Question
For a batch of 16 images with 32 channels and 64×64 spatial resolution, over how many total values does InstanceNorm compute the mean and variance for one specific (example, channel) pair?
