Instance norm normalises each sample's each channel over its spatial positions only, so the style (contrast) of one image doesn't depend on others in the batch.
Implement instance_norm(x, gamma, beta, eps=1e-5) for x of shape (N, C, H, W) and per-channel gamma, beta of shape (C,). Use the biased variance.