2nd/10_Wiki/Topics/Architecture/Variational-Autoencoders-VAE.md

---
id: wiki-2026-0508-variational-autoencoders-vae
title: Variational Autoencoders (VAE)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [VAE, Variational Autoencoder, β-VAE]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [generative-model, deep-learning, latent-variable, variational-inference]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: PyTorch 2.x
---

# Variational Autoencoders (VAE)

## 매 한 줄
> **"매 encoder 가 input 의 latent distribution (μ, σ) 의 produce → reparameterization trick 으로 sample → decoder 의 reconstruct. 매 ELBO = reconstruction loss + KL(q(z|x) || p(z))"**. 매 Kingma & Welling 2013 (Auto-Encoding Variational Bayes). 매 2026 의 modern role: standalone generation 의 X (diffusion 의 우위) BUT 매 Stable Diffusion / FLUX / Sora 의 latent space 의 backbone — 매 image 의 8× downsampled latent 의 work.

## 매 핵심

### 매 수학 (ELBO)
- **Goal**: maximize log p(x). 매 intractable.
- **Trick**: variational posterior q_φ(z|x) ≈ p(z|x). 매 ELBO 의 lower bound.
- **ELBO** = E_{z~q}[log p_θ(x|z)] − D_KL(q_φ(z|x) || p(z))
  - 1번 term: reconstruction (decoder).
  - 2번 term: regularize latent 의 prior (보통 N(0,I)) 에 가깝게.
- **Reparameterization**: z = μ + σ ⊙ ε, ε~N(0,I) — 매 backprop through stochastic sampling.

### 매 vs 다른 generative
- **GAN**: sharp, no likelihood, mode collapse. VAE: blurry, likelihood, stable training.
- **Diffusion**: state-of-art quality. VAE: faster inference (single forward).
- **2026 dominant role**: latent diffusion 의 frontend — 매 VAE 가 pixel space → latent space 압축, diffusion 이 latent 의 denoise.

### 매 변종
- **β-VAE**: KL term 에 β 곱 → β>1 의 disentangled latent.
- **VQ-VAE**: continuous latent → discrete codebook (Vector Quantization). 매 DALL-E, Sora 의 핵심.
- **Hierarchical VAE / NVAE**: multi-scale latents.
- **Conditional VAE (CVAE)**: conditional generation p(x|c).

### 매 응용
1. Latent diffusion (Stable Diffusion / FLUX / Sora) — 매 8×8 patch → 4-ch latent.
2. Anomaly detection — high reconstruction error = anomaly.
3. Molecular generation — 매 chemistry latent space exploration.

## 💻 패턴

### Vanilla VAE (PyTorch 2.x)
```python
import torch
import torch.nn as nn
import torch.nn.functional as F

class VAE(nn.Module):
    def __init__(self, in_dim=784, hidden=400, z_dim=20):
        super().__init__()
        self.fc1 = nn.Linear(in_dim, hidden)
        self.fc_mu = nn.Linear(hidden, z_dim)
        self.fc_logvar = nn.Linear(hidden, z_dim)
        self.fc2 = nn.Linear(z_dim, hidden)
        self.fc3 = nn.Linear(hidden, in_dim)

    def encode(self, x):
        h = F.relu(self.fc1(x))
        return self.fc_mu(h), self.fc_logvar(h)

    def reparameterize(self, mu, logvar):
        std = torch.exp(0.5 * logvar)
        eps = torch.randn_like(std)
        return mu + eps * std

    def decode(self, z):
        return torch.sigmoid(self.fc3(F.relu(self.fc2(z))))

    def forward(self, x):
        mu, logvar = self.encode(x)
        z = self.reparameterize(mu, logvar)
        return self.decode(z), mu, logvar

def vae_loss(recon, x, mu, logvar):
    bce = F.binary_cross_entropy(recon, x, reduction='sum')
    kld = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
    return bce + kld
```

### Training loop
```python
model = VAE().cuda()
opt = torch.optim.AdamW(model.parameters(), lr=1e-3)
for epoch in range(50):
    for x, _ in loader:
        x = x.view(-1, 784).cuda()
        recon, mu, logvar = model(x)
        loss = vae_loss(recon, x, mu, logvar)
        opt.zero_grad(); loss.backward(); opt.step()
```

### β-VAE (disentanglement)
```python
def beta_vae_loss(recon, x, mu, logvar, beta=4.0):
    bce = F.binary_cross_entropy(recon, x, reduction='sum')
    kld = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
    return bce + beta * kld
```

### VQ-VAE (vector quantization)
```python
class VectorQuantizer(nn.Module):
    def __init__(self, num_embeddings=512, embedding_dim=64, commitment=0.25):
        super().__init__()
        self.embed = nn.Embedding(num_embeddings, embedding_dim)
        self.embed.weight.data.uniform_(-1.0 / num_embeddings, 1.0 / num_embeddings)
        self.commitment = commitment

    def forward(self, z_e):  # z_e: (B, C, H, W)
        z_e_perm = z_e.permute(0, 2, 3, 1).contiguous()
        flat = z_e_perm.view(-1, z_e_perm.size(-1))
        # Nearest codebook vector
        d = (flat.pow(2).sum(1, keepdim=True)
             - 2 * flat @ self.embed.weight.t()
             + self.embed.weight.pow(2).sum(1))
        idx = d.argmin(1)
        z_q = self.embed(idx).view(z_e_perm.shape).permute(0, 3, 1, 2)
        # Straight-through estimator
        loss = F.mse_loss(z_q.detach(), z_e) + self.commitment * F.mse_loss(z_q, z_e.detach())
        z_q = z_e + (z_q - z_e).detach()
        return z_q, loss, idx
```

### Sample / generate
```python
model.eval()
with torch.no_grad():
    z = torch.randn(64, 20).cuda()
    samples = model.decode(z).view(-1, 1, 28, 28).cpu()
```

### Latent diffusion VAE (SD-style — using diffusers)
```python
from diffusers import AutoencoderKL
import torch

vae = AutoencoderKL.from_pretrained('stabilityai/sd-vae-ft-mse').cuda()
# Encode 512x512 image → 4-ch 64x64 latent
img = torch.randn(1, 3, 512, 512).cuda()
latent = vae.encode(img).latent_dist.sample() * vae.config.scaling_factor
# Diffusion happens in latent space, then decode
recon = vae.decode(latent / vae.config.scaling_factor).sample
```

## 매 결정 기준
| 목적 | Choice |
|---|---|
| 2026 SOTA 이미지 생성 | Diffusion (FLUX, Stable Diffusion 3.5) — 매 VAE 의 frontend 만 |
| Disentangled representation 연구 | β-VAE |
| Discrete latent (LLM tokenize 유사) | VQ-VAE / VQ-GAN |
| Anomaly detection | Vanilla VAE — reconstruction error |
| Latent diffusion 학습 | Pre-trained KL-regularized VAE (e.g. SD VAE) reuse |
| Molecular / structured generation | VAE (continuous latent) — 매 still competitive |

**기본값**: 매 image generation 의 directly 의 X — 매 latent diffusion 안 의 VAE 로 사용. 매 disentanglement / anomaly 의 standalone VAE.

## 🔗 Graph
- 부모: [[Variational Inference]]
- 변형: [[β-VAE]]
- 응용: [[Stable Diffusion]] · [[FLUX]]
- Adjacent: [[Diffusion Models]] · [[Generative-Adversarial-Networks|GAN]]

## 🤖 LLM 활용
**언제**: 매 latent diffusion 의 VAE component 설명, 매 anomaly detection baseline 작성, 매 ELBO 수학 의 derivation, 매 reparameterization trick 의 implementation.
**언제 X**: 매 standalone SOTA image generation (diffusion 우선), 매 sharp output 필수 (GAN/diffusion).

## ❌ 안티패턴
- **Posterior collapse**: q(z|x) → p(z) 의 무시 → KL=0, decoder 의 z 의 ignore. 매 KL annealing / β scheduling 필요.
- **Pixel-space VAE 의 high-res 직접**: 매 blurry, 매 8× downsample latent + diffusion 으로 decouple.
- **σ 의 직접 output**: 매 negative 가능. 매 logvar 의 output → σ = exp(0.5 * logvar).
- **KL 의 mean reduction**: 매 batch mean 의 reconstruction 의 sum 과 mismatch — 매 두 term 의 same reduction.

## 🧪 검증 / 중복
- Verified (Kingma & Welling 2013 ICLR, Stable Diffusion paper, NVIDIA NVAE, DeepMind β-VAE).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full VAE with ELBO, β-VAE, VQ-VAE, latent diffusion role, 6 patterns |