---
id: wiki-2026-0508-style-transfer
title: Style Transfer
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Neural Style Transfer, NST, Style Transfer in AI]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [style-transfer, neural-network, image-generation, diffusion]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pytorch
---

# Style Transfer

## 매 한 줄
> **"매 content + style separation의 art"**. Gatys et al. (2015) 가 VGG feature space 의 Gram matrix 로 style 추출 → 매 content image 에 transfer 의 seminal work. 2026 의 modern state 는 diffusion-based (IP-Adapter, ControlNet style) + Midjourney --sref 의 mainstream.

## 매 핵심

### 매 origin (Gatys 2015)
- VGG-19 의 conv layer activation 의 feature representation.
- **Content loss**: 매 high-level layer (conv4_2) 의 feature MSE.
- **Style loss**: 매 multiple layer 의 Gram matrix (feature correlation) MSE.
- Optimization-based — 매 image pixel 자체 의 gradient descent (slow, ~minutes per image).

### 매 evolution
- **Fast NST** (Johnson 2016): feedforward network 의 single forward pass.
- **AdaIN** (Huang 2017): Adaptive Instance Normalization — 매 arbitrary style 의 real-time.
- **Diffusion-based** (2023+): IP-Adapter, ControlNet — 매 prompt + reference 의 zero-shot.

### 매 응용
1. Artistic image generation (Prisma, DeepArt — 매 historical).
2. Midjourney --sref / --cref — 매 mainstream creative tool.
3. Video stylization (Runway, Kaiber).
4. Domain adaptation (synthetic → real).

## 💻 패턴

### Gram matrix (style representation)
```python
import torch
import torch.nn as nn

def gram_matrix(features):
    b, c, h, w = features.shape
    feat = features.view(b, c, h * w)
    gram = torch.bmm(feat, feat.transpose(1, 2))
    return gram / (c * h * w)
```

### AdaIN
```python
def adain(content_feat, style_feat, eps=1e-5):
    c_mean = content_feat.mean(dim=[2, 3], keepdim=True)
    c_std = content_feat.std(dim=[2, 3], keepdim=True) + eps
    s_mean = style_feat.mean(dim=[2, 3], keepdim=True)
    s_std = style_feat.std(dim=[2, 3], keepdim=True) + eps
    normalized = (content_feat - c_mean) / c_std
    return normalized * s_std + s_mean
```

### Gatys optimization (full)
```python
import torch.optim as optim
from torchvision.models import vgg19

vgg = vgg19(pretrained=True).features.eval().cuda()
target = content_img.clone().requires_grad_(True)
optimizer = optim.LBFGS([target])

def closure():
    optimizer.zero_grad()
    feats = extract_features(target, vgg)
    c_loss = F.mse_loss(feats['content'], content_feats['content'])
    s_loss = sum(F.mse_loss(gram_matrix(feats[l]), gram_matrix(style_feats[l]))
                 for l in style_layers)
    loss = c_loss + 1e6 * s_loss
    loss.backward()
    return loss

for _ in range(300):
    optimizer.step(closure)
```

### IP-Adapter (diffusion-based, 2024+)
```python
from diffusers import StableDiffusionPipeline
from ip_adapter import IPAdapter

pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5").to("cuda")
ip_model = IPAdapter(pipe, "h94/IP-Adapter", "models/ip-adapter_sd15.bin", "cuda")

style_ref = Image.open("vangogh.jpg")
images = ip_model.generate(pil_image=style_ref, prompt="a cat", num_samples=4, scale=0.7)
```

### Midjourney --sref (2026 mainstream)
```
/imagine prompt: a serene lake at sunset --sref https://example.com/style.jpg --sw 100
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 일회성 art experiment | Gatys (구현 simple, slow ok) |
| 매 real-time / video | AdaIN, fast NST |
| 매 production creative | IP-Adapter + SDXL / FLUX |
| 매 non-coder creative | Midjourney --sref |
| 매 controllable structure + style | ControlNet + IP-Adapter combo |

**기본값**: 매 2026 의 IP-Adapter (open) 또는 Midjourney --sref (closed).

## 🔗 Graph
- 부모: [[Generative-AI]] · [[Computer Vision|Computer-Vision]]
- 변형: [[Style_Reference_(--sref)]] · [[ControlNet]] · [[IP-Adapter]]
- 응용: [[AI 이미지 생성 (AI Image Generation)|Image-Generation]]
- Adjacent: [[Diffusion-Models]]

## 🤖 LLM 활용
**언제**: 매 creative pipeline 의 style consistency, brand asset variant 생성, mood board 의 visual exploration.
**언제 X**: 매 photo retouching (use Lightroom), 매 strict color grading (use LUTs), 매 face identity preservation 의 unstable.

## ❌ 안티패턴
- **Style weight 무한 증가**: 매 content 가 사라짐. balance 필수 (1e6 typical).
- **Single VGG layer**: 매 multi-scale style 의 lost. 매 multiple layer aggregate.
- **Diffusion 의 prompt 무시**: IP-Adapter scale 너무 높으면 prompt 의 무시. scale 0.5-0.8 sweet spot.

## 🧪 검증 / 중복
- Verified (Gatys et al. 2015 "A Neural Algorithm of Artistic Style"; Huang & Belongie 2017 AdaIN; Ye et al. 2023 IP-Adapter).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Gatys → AdaIN → diffusion (IP-Adapter, --sref) coverage |