---
id: wiki-2026-0508-generative-adversarial-networks
title: Generative Adversarial Networks (GAN)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [GAN, generative adversarial networks, StyleGAN, CycleGAN, Goodfellow, Wasserstein GAN]
duplicate_of: none
source_trust_level: A
confidence_score: 0.97
verification_status: applied
tags: [deep-learning, gan, generative, goodfellow, stylegan, image-generation]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: PyTorch / TensorFlow
---

# Generative Adversarial Networks (GAN)

## 매 한 줄
> **"매 generator 의 의 의 fake 의 의 의, discriminator 의 의 의 detect — 매 minimax game"**. Goodfellow 2014. 매 image gen 의 dominant (2014-2022) → 매 diffusion 의 의 의 displace. 매 modern: 매 StyleGAN3, 매 CycleGAN, 매 GAN-inversion.

## 매 핵심

### 매 model
- **Generator G**: 매 noise → fake.
- **Discriminator D**: 매 real or fake.
- **Loss**: G 의 의 fool D, D 의 의 catch.

### 매 famous variants
- **DCGAN** (Radford 2015): 매 conv-based.
- **WGAN** (Arjovsky 2017): 매 Wasserstein distance.
- **WGAN-GP**: 매 gradient penalty.
- **StyleGAN** v1/v2/v3 (Karras): 매 face quality.
- **CycleGAN**: 매 unpaired image translation.
- **Pix2Pix**: 매 paired translation.
- **BigGAN**: 매 class-conditional large.
- **GAN inversion**: 매 image → latent.

### 매 modern context (2024+)
- **Diffusion** dominate text-to-image.
- **GAN niche**: 매 fast inference, 매 specific style.
- **GAN inversion** for editing.
- **StyleGAN** still SOTA for faces.

### 매 응용
1. **Image gen** (faces).
2. **Style transfer**.
3. **Super-resolution** (ESRGAN).
4. **Image-to-image** (CycleGAN).
5. **Data augmentation**.
6. **Anomaly detection** (AnoGAN).

## 💻 패턴

### DCGAN (PyTorch)
```python
import torch
import torch.nn as nn

class Generator(nn.Module):
    def __init__(self, latent=100, img_dim=64):
        super().__init__()
        self.net = nn.Sequential(
            nn.ConvTranspose2d(latent, 512, 4, 1, 0), nn.BatchNorm2d(512), nn.ReLU(),
            nn.ConvTranspose2d(512, 256, 4, 2, 1), nn.BatchNorm2d(256), nn.ReLU(),
            nn.ConvTranspose2d(256, 128, 4, 2, 1), nn.BatchNorm2d(128), nn.ReLU(),
            nn.ConvTranspose2d(128, 64, 4, 2, 1), nn.BatchNorm2d(64), nn.ReLU(),
            nn.ConvTranspose2d(64, 3, 4, 2, 1), nn.Tanh(),
        )
    
    def forward(self, z):
        return self.net(z.unsqueeze(-1).unsqueeze(-1))

class Discriminator(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(3, 64, 4, 2, 1), nn.LeakyReLU(0.2),
            nn.Conv2d(64, 128, 4, 2, 1), nn.BatchNorm2d(128), nn.LeakyReLU(0.2),
            nn.Conv2d(128, 256, 4, 2, 1), nn.BatchNorm2d(256), nn.LeakyReLU(0.2),
            nn.Conv2d(256, 1, 4, 1, 0), nn.Sigmoid(),
        )
    
    def forward(self, x):
        return self.net(x).view(-1)
```

### Train loop
```python
G, D = Generator(), Discriminator()
opt_G = torch.optim.Adam(G.parameters(), lr=2e-4, betas=(0.5, 0.999))
opt_D = torch.optim.Adam(D.parameters(), lr=2e-4, betas=(0.5, 0.999))
bce = nn.BCELoss()

for batch in dataloader:
    real = batch.cuda()
    bs = real.size(0)
    z = torch.randn(bs, 100).cuda()
    fake = G(z)
    
    # 매 D
    opt_D.zero_grad()
    d_real = D(real)
    d_fake = D(fake.detach())
    d_loss = bce(d_real, torch.ones(bs).cuda()) + bce(d_fake, torch.zeros(bs).cuda())
    d_loss.backward(); opt_D.step()
    
    # 매 G
    opt_G.zero_grad()
    d_fake_g = D(fake)
    g_loss = bce(d_fake_g, torch.ones(bs).cuda())
    g_loss.backward(); opt_G.step()
```

### WGAN-GP loss
```python
def gradient_penalty(D, real, fake, device):
    bs = real.size(0)
    alpha = torch.rand(bs, 1, 1, 1, device=device)
    interp = alpha * real + (1 - alpha) * fake
    interp.requires_grad_(True)
    d_interp = D(interp)
    grads = torch.autograd.grad(d_interp.sum(), interp, create_graph=True)[0]
    return ((grads.norm(2, dim=[1,2,3]) - 1) ** 2).mean()

# 매 D loss
d_loss = D(fake).mean() - D(real).mean() + 10 * gradient_penalty(D, real, fake, device)
```

### CycleGAN (unpaired translation)
```python
class CycleGAN:
    def __init__(self):
        self.G_AB, self.G_BA = Generator(), Generator()
        self.D_A, self.D_B = Discriminator(), Discriminator()
    
    def cycle_loss(self, real_A, real_B):
        fake_B = self.G_AB(real_A)
        rec_A = self.G_BA(fake_B)
        cycle_A = (rec_A - real_A).abs().mean()
        
        fake_A = self.G_BA(real_B)
        rec_B = self.G_AB(fake_A)
        cycle_B = (rec_B - real_B).abs().mean()
        
        return cycle_A + cycle_B
```

### StyleGAN (style modulation)
```python
class StyleBlock(nn.Module):
    def __init__(self, in_ch, out_ch, style_dim=512):
        super().__init__()
        self.conv = nn.Conv2d(in_ch, out_ch, 3, padding=1)
        self.style_proj = nn.Linear(style_dim, in_ch)
    
    def forward(self, x, style):
        scale = self.style_proj(style).unsqueeze(-1).unsqueeze(-1)
        x = x * scale
        return self.conv(x)
```

### Spectral normalization
```python
from torch.nn.utils import spectral_norm
class SNDiscriminator(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = spectral_norm(nn.Conv2d(3, 64, 4, 2, 1))
        # ...
```

### Mode collapse detection
```python
def diversity_check(generator, n=1000):
    z = torch.randn(n, 100).cuda()
    fake = generator(z)
    # 매 LPIPS pairwise
    distances = []
    for i in range(min(100, n)):
        for j in range(i+1, min(100, n)):
            distances.append(lpips(fake[i:i+1], fake[j:j+1]).item())
    return np.mean(distances)  # 매 low = mode collapse
```

### FID (eval)
```python
from torchmetrics.image.fid import FrechetInceptionDistance
fid = FrechetInceptionDistance().cuda()
fid.update(real_imgs, real=True)
fid.update(fake_imgs, real=False)
print(fid.compute())  # 매 lower = better
```

### GAN inversion (project image → latent)
```python
def gan_invert(target_image, G, n_iter=1000):
    z = torch.randn(1, 100, requires_grad=True, device='cuda')
    optim = torch.optim.Adam([z], lr=0.01)
    for _ in range(n_iter):
        gen = G(z)
        loss = ((gen - target_image) ** 2).mean() + lpips_loss(gen, target_image)
        optim.zero_grad(); loss.backward(); optim.step()
    return z
```

### Conditional (class-conditional)
```python
class CondG(nn.Module):
    def __init__(self, n_classes):
        super().__init__()
        self.embed = nn.Embedding(n_classes, 100)
        self.gen = Generator()
    
    def forward(self, z, label):
        return self.gen(z + self.embed(label))
```

### ESRGAN (super-resolution)
```python
# 매 conceptual
class RRDB(nn.Module):  # 매 residual-in-residual dense block
    pass

class ESRGAN(nn.Module):
    def __init__(self):
        super().__init__()
        self.body = nn.ModuleList([RRDB() for _ in range(23)])
        self.upsample = nn.Sequential(
            nn.Conv2d(64, 64*4, 3, padding=1), nn.PixelShuffle(2),
            nn.Conv2d(64, 3, 3, padding=1),
        )
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Modern image gen | Diffusion (not GAN) |
| Face generation | StyleGAN3 |
| Unpaired translation | CycleGAN |
| Paired translation | Pix2Pix |
| Super-resolution | ESRGAN |
| Fast inference | GAN > Diffusion |
| Editing | GAN inversion |

**기본값**: 매 modern = diffusion. 매 face = StyleGAN3. 매 niche translation = CycleGAN. 매 SR = ESRGAN. 매 always FID + diversity check.

## 🔗 Graph
- 부모: [[Deep Learning]] · [[Generative-AI|Generative-Models]]
- 변형: [[StyleGAN]] · [[CycleGAN]] · [[Pix2Pix]]
- 응용: [[Data-Augmentation]]
- Adjacent: [[Diffusion-Models]] · [[VAE]] · [[Generative-AI]]

## 🤖 LLM 활용
**언제**: 매 fast image gen. 매 unpaired translation. 매 face.
**언제 X**: 매 text-to-image (use diffusion).

## ❌ 안티패턴
- **Mode collapse 의 ignore**: 매 limited diversity.
- **No spectral norm**: 매 unstable D.
- **Imbalanced D-G**: 매 collapse.
- **No FID**: 매 quality 의 invisible.
- **GAN for text-to-image**: 매 diffusion 이 better.

## 🧪 검증 / 중복
- Verified (Goodfellow 2014, Karras StyleGAN, Zhu CycleGAN).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-26 | Auto |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — DCGAN/WGAN/Style/Cycle + 매 train / inversion / FID code |