Neural Style Transfer with Adaptive Normalization

Problem

Classical neural style transfer (Gatys et al.) is iterative and takes minutes per image, making it impractical for interactive use or video. We needed a feed-forward approach that generalises to arbitrary style images without retraining.

Approach

We adopted Adaptive Instance Normalization (AdaIN) to align the mean and variance of content features with those of the style image. A lightweight decoder reconstructs the stylised image from the normalised feature space in a single forward pass.

Results

Metric	Value
Inference time (512×512)	18ms
Hardware	RTX 3080
Style generalisation	Arbitrary (no retraining)

Stylised video runs at 55 FPS with temporal consistency filtering applied.

---

This is a placeholder project entry. Replace with real content in Phase 8.