2nd/10_Wiki/Topics/AI_and_ML/Deep-Convolutional-GANs.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-dcgan	DCGAN (Deep Convolutional GAN)	10_Wiki/Topics	verified	self	DCGAN GAN generative adversarial network StyleGAN CycleGAN Pix2Pix	none	A	0.88	applied	gan dcgan generative-models deep-learning image-generation stylegan cyclegan history		2026-05-10	pending	language framework Python PyTorch

DCGAN

매 한 줄

"매 GAN 의 first 의 stable architecture" (Radford 2015). 매 stride conv + 매 batch norm + 매 specific activation. 매 generative AI 의 grandparent. 매 modern: 매 Diffusion 의 superseded 가, 매 fast inference / GAN-based super-res / image-to-image 의 still relevant.

매 핵심

매 GAN basics (Goodfellow 2014)

• Generator: 매 noise → 매 image.
• Discriminator: 매 real vs fake.
• Min-max game.

DCGAN (2015) 의 contribution

1. Strided conv (no pooling).
2. BatchNorm in both G and D.
3. No fully-connected hidden layer.
4. ReLU in G (Tanh output).
5. LeakyReLU in D.

매 famous GAN evolution

• DCGAN (2015): 매 stable.
• WGAN (2017): 매 Wasserstein loss, 매 mode collapse 의 mitigate.
• Pix2Pix (2017): 매 image-to-image.
• CycleGAN (2017): 매 unpaired.
• StyleGAN (2018-2021): 매 face quality SOTA.
• BigGAN (2018): 매 large-scale.
• GigaGAN (2023): 매 text-to-image GAN.

매 mode collapse

• 매 G 의 매 limited variety 의 generate.
• 매 mitigation: 매 minibatch discrimination, 매 spectral norm, 매 WGAN.

매 evaluation

• FID (Fréchet Inception Distance): 매 generated vs real 의 distance.
• IS (Inception Score).
• Precision / Recall (Kynkäänniemi 2019).

매 modern relevance

• Super-resolution (ESRGAN, Real-ESRGAN).
• Image-to-image (CycleGAN 의 still useful).
• Domain adaptation (sim2real).
• Implicit EBM.
• Fast generation (vs slow diffusion).

매 vs Diffusion

측면	GAN	Diffusion
Quality	High (StyleGAN)	Highest
Diversity	Mode collapse risk	High
Training stability	Tricky	Stable
Inference speed	Fast (single step)	Slow (multi-step)
Conditioning	Hard	Easy (CLIP)

💻 패턴

DCGAN (PyTorch)

import torch.nn as nn

class Generator(nn.Module):
    def __init__(self, nz=100, ngf=64, nc=3):
        super().__init__()
        self.main = nn.Sequential(
            # 매 input: nz × 1 × 1
            nn.ConvTranspose2d(nz, ngf*8, 4, 1, 0, bias=False),
            nn.BatchNorm2d(ngf*8),
            nn.ReLU(True),
            # 매 ngf*8 × 4 × 4
            nn.ConvTranspose2d(ngf*8, ngf*4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ngf*4),
            nn.ReLU(True),
            # 매 ... → 매 nc × 64 × 64
            nn.ConvTranspose2d(ngf, nc, 4, 2, 1, bias=False),
            nn.Tanh(),
        )
    
    def forward(self, z):
        return self.main(z)

class Discriminator(nn.Module):
    def __init__(self, ndf=64, nc=3):
        super().__init__()
        self.main = nn.Sequential(
            nn.Conv2d(nc, ndf, 4, 2, 1, bias=False),
            nn.LeakyReLU(0.2, True),
            nn.Conv2d(ndf, ndf*2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ndf*2),
            nn.LeakyReLU(0.2, True),
            # 매 ... 
            nn.Conv2d(ndf*8, 1, 4, 1, 0, bias=False),
            nn.Sigmoid(),
        )
    
    def forward(self, x):
        return self.main(x).view(-1)

Training loop

G = Generator().to('cuda')
D = Discriminator().to('cuda')
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))
criterion = nn.BCELoss()

for real, _ in loader:
    real = real.to('cuda')
    bs = real.size(0)
    
    # 매 D step
    opt_d.zero_grad()
    d_real = D(real)
    loss_real = criterion(d_real, torch.ones(bs).to('cuda'))
    
    z = torch.randn(bs, 100, 1, 1).to('cuda')
    fake = G(z)
    d_fake = D(fake.detach())
    loss_fake = criterion(d_fake, torch.zeros(bs).to('cuda'))
    
    (loss_real + loss_fake).backward()
    opt_d.step()
    
    # 매 G step
    opt_g.zero_grad()
    d_fake = D(fake)
    loss_g = criterion(d_fake, torch.ones(bs).to('cuda'))
    loss_g.backward()
    opt_g.step()

WGAN-GP (modern stable)

def gradient_penalty(D, real, fake):
    bs = real.size(0)
    alpha = torch.rand(bs, 1, 1, 1).to(real.device)
    interp = alpha * real + (1 - alpha) * fake
    interp.requires_grad_()
    
    d_interp = D(interp)
    grads = torch.autograd.grad(d_interp.sum(), interp, create_graph=True)[0]
    grad_norm = grads.view(bs, -1).norm(2, dim=1)
    return ((grad_norm - 1) ** 2).mean()

# 매 D step
loss_d = D(fake).mean() - D(real).mean() + 10 * gradient_penalty(D, real, fake)

Real-ESRGAN (super-resolution, modern application)

from realesrgan import RealESRGAN
from PIL import Image

model = RealESRGAN(device='cuda', scale=4)
model.load_weights('weights/RealESRGAN_x4.pth', download=True)

img = Image.open('low_res.jpg')
sr_img = model.predict(img)
sr_img.save('high_res.jpg')

CycleGAN (unpaired image-to-image)

# 매 매 horse → 매 zebra (no pair)
# 매 G_ab: 매 A → B, 매 G_ba: 매 B → A
# 매 cycle loss: 매 G_ba(G_ab(a)) ≈ a

StyleGAN inversion (modern)

from stylegan2_pytorch import Trainer

# 매 매 image 의 latent z 의 find
def invert(image, generator, n_iters=1000):
    z = torch.randn(1, 512, requires_grad=True)
    for _ in range(n_iters):
        gen = generator(z)
        loss = (gen - image).pow(2).mean()
        loss.backward()
        z.data -= 0.01 * z.grad.data
        z.grad.zero_()
    return z

FID evaluation

from pytorch_fid import fid_score

fid = fid_score.calculate_fid_given_paths(
    ['./real_images/', './fake_images/'],
    batch_size=50,
    device='cuda',
    dims=2048,
)
print(f'FID: {fid:.2f}')  # 매 lower 의 better

매 결정 기준

응용	Method
Photoreal generation	Diffusion (SDXL, Flux)
Face	StyleGAN3
Super-resolution	Real-ESRGAN
Domain adapt (unpaired)	CycleGAN
Sim2Real	CycleGAN / paired
Fast inference	GAN > Diffusion
Quality + control	Diffusion + ControlNet

기본값: 매 modern generation = Diffusion. 매 GAN = 매 SR / I2I 의 still.

🔗 Graph

• 부모: Generative-AI · Deep Learning
• 변형: Generative-Adversarial-Networks · StyleGAN · Pix2Pix · CycleGAN
• 응용: Domain-Adaptation · CV_Synthesis
• Adjacent: Diffusion-Models · Auto-Encoding · Stable-Diffusion · Deepfake-Technology

🤖 LLM 활용

언제: 매 GAN history. 매 fast generation. 매 super-res. 매 image-to-image. 언제 X: 매 highest quality (use diffusion).

❌ 안티패턴

• DCGAN 의 production 의 force: 매 modern 의 diffusion 더 좋음.
• No mode collapse check: 매 single output.
• WGAN-GP 없 의 unstable: 매 training fail.
• FID 만 의 trust: 매 다른 metric 도.

🧪 검증 / 중복

• Verified (Goodfellow GAN, Radford DCGAN, Karras StyleGAN, Real-ESRGAN).
• 신뢰도 A.
• Related: Diffusion-Models · Auto-Encoding · Stable-Diffusion · Deepfake-Technology · CV_Synthesis · Bioenergetics (model collapse).

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — DCGAN architecture + GAN evolution + 매 PyTorch / WGAN-GP / Real-ESRGAN code