AdaIN (Adaptive Instance Normalization) is a style transfer technique that transfers style by matching the mean and variance of content feature maps to those of style feature maps — enabling real-time arbitrary style transfer with a single forward pass.

How Does AdaIN Work?

• Formula: $AdaIN(x, y) = sigma(y) cdot frac{x - mu(x)}{sigma(x)} + mu(y)$
• Process: Normalize content features $x$ to zero mean/unit variance (InstanceNorm), then scale and shift using style features' statistics $sigma(y), mu(y)$.
• Single Pass: No iterative optimization needed (unlike Gatys et al. style transfer).
• Paper: Huang & Belongie (2017).

Why It Matters

• Real-Time: Arbitrary style transfer at inference speed — any style, any content, one forward pass.
• StyleGAN: AdaIN (and its evolution, style modulation) is the core mechanism of the StyleGAN architecture.
• Foundation: The insight that style information is captured in feature statistics (mean + variance) is profound.

AdaIN is the statistics swap that enables neural style transfer — exchanging mean and variance to paint any content in any style in real time.