style loss

**Style loss** is a **perceptual loss that measures texture and style similarity via Gram matrix feature correlations** — capturing texture patterns, color distributions, and artistic style by comparing second-order feature statistics rather than spatial structure, enabling neural style transfer and texture synthesis without preserving specific object layouts. **Mathematical Foundation** Gram matrix G of feature map F: ``` G_ij = Σ_spatial F_i * F_j (correlation between channels) ``` Style loss measures feature correlation differences, capturing texture without spatial structure. **Key Components** - **Gram Matrices**: Encode texture statistics across channels - **Multi-scale**: Apply across VGG layers (conv1-5) for diverse style - **Invariant**: Agnostic to spatial arrangement — captures style essence - **Perceptual**: More meaningful than pixel-wise Euclidean distance **Applications** Neural style transfer combining content and style losses, texture synthesis, artistic rendering, photo-realistic style adaptation. Style loss captures **texture and artistic essence** — separating style from structure for transfer tasks.