## TL;DR
Neural style transfer applies the artistic style of one image (e.g., Van Gogh painting) to the content of another (a photograph), creating new artwork. From Gatys's seminal 2015 paper to IP-Adapter diffusion models, the field has evolved from minute-long optimization to real-time fine-grained style control.

## Core Explanation
Gatys et al. (2015): given a content image C and style image S, optimize a generated image G to minimize content loss (||VGG_features(G) - VGG_features(C)|| at layer conv4_2) + style loss (difference in Gram matrices at layers conv1_1 through conv5_1). Gram matrix captures texture correlations. Evolution: (1) Slow optimization (2015) -- minutes per image; (2) Feed-forward (2016) -- train a generator network per style, real-time inference (Johnson et al.); (3) Arbitrary style (2017-2019) -- Adaptive Instance Normalization (AdaIN), single model for any style; (4) Diffusion-based (2023-present) -- ControlNet conditions generation on structural cues, IP-Adapter injects style via image embedding. Diffusion methods separate content structure (edge map, depth map) from style conditions.

## Detailed Analysis
AdaIN (Huang & Belongie, ICCV 2017): aligns channel-wise mean and variance of content features to match style features. Simple, fast, effective. ControlNet (2023): copies Stable Diffusion UNet encoder, fine-tunes on task-specific conditioning (canny edges, depth maps, pose) while freezing base model. For style transfer: edge map from content + style image as IP-Adapter conditioning. IP-Adapter (Tencent, 2023): decoupled cross-attention injecting style features into diffusion model. Only 22M trainable parameters. DragGAN (2023) and Inpaint4Drag (ICCV 2025): interactive point-based manipulation of images. arxiv 2025 survey: diffusion-based methods solve the content-style trade-off better than prior approaches -- high style fidelity without content distortion. Evaluation: CLIP score (text-image alignment), user preference studies, artist panel evaluation. Key challenge: no objective metric for "good style transfer" -- evaluation remains inherently subjective.