GANs from DCGAN to StyleGAN · Ingénierie IA

Losses, DCGAN code, Pix2Pix, CycleGAN and StyleGAN. Defaults, failure signatures and fixes.

StyleGAN generator

Nodes and connections

Node	Role	Kind
z	N(0, I), 512 dims	store
Mapping network	8-layer MLP	model
w in W	roughly disentangled	store
Learned constant	4×4×512	store
Synthesis block	upsample, conv, AdaIN, noise	gpu
Image	4 up to 1024	client
Per-layer noise	learned per-channel scale	process

From	To	Connection
z	Mapping network
Mapping network	w in W
w in W	Synthesis block	AdaIN at every block
Learned constant	Synthesis block
Per-layer noise	Synthesis block
Synthesis block	Image	resolution doubles

Sources

phases/08-generative-ai/05-stylegan/docs/en.md

One training step

Sample draw a real batch and noise z from N(0, I).
D step score real images against 1 and G(z).detach() against 0.
G step make fresh fakes. Score D(G(z)) against 1.
Repeat one step of D, then one step of G, every batch.
Probe every 100 steps, count samples per mode. Warn below 50.
Inspect look at 64 samples at the end of every epoch.
Equilibrium p_G = p_data and D outputs 0.5 everywhere.

Sources

Losses and formulas

Minimax: \min_G \max_D \; \mathbb{E}_x[\log D(x)] + \mathbb{E}_z[\log(1 - D(G(z)))]
D loss: binary cross-entropy with real = 1 and fake = 0.
G loss: -\log D(G(z)). This is the non-saturating form.
WGAN: \text{loss}_D = \mathbb{E}[D(\text{fake})] - \mathbb{E}[D(\text{real})], \text{loss}_G = -\mathbb{E}[D(\text{fake})]
Pix2Pix G: -\log D(x, G(x)) + \lambda \cdot \lVert y - G(x) \rVert_1
Cycle: F(G(x)) \approx x and G(F(y)) \approx y.
AdaIN: y_{\text{scale}} \cdot \frac{x - \mathrm{mean}(x)}{\mathrm{std}(x)} + y_{\text{bias}}
Truncation: w' = \hat{w} + \psi \cdot (w - \hat{w}). \psi < 1 trades diversity for quality.

Sources

DCGAN train_step

opt_d.zero_grad()

d_real = D(real)logits, sigmoid only in the loss

d_fake = D(G(z).detach())no gradient flows into G

loss_d = (F.binary_cross_entropy_with_logits(d_real, torch.ones_like(d_real))

+ F.binary_cross_entropy_with_logits(d_fake, torch.zeros_like(d_fake)))

loss_d.backward()

opt_d.step()

opt_g.zero_grad()

d_fake = D(G(z))second pass keeps the graph

loss_g = F.binary_cross_entropy_with_logits(d_fake, torch.ones_like(d_fake))

loss_g.backward()non-saturating G loss

opt_g.step()

Sources

Tasks in code

Task	Code
Build the generator	`Generator(z_dim=64, img_channels=3, feat=32)`
Wrap a D conv in SN	`spectral_norm(c)`
Make the optimizers	`Adam(lr=2e-4, betas=(0.5, 0.999))`
Sample in eval mode	`sample(G, n=8, z_dim=z_dim, device=device)`
Count parameters	`sum(p.numel() for p in G.parameters())`
Condition G or D	`inp = z + one_hot(c, num_classes)`
Map z to w	`w = mapping(z, mapping_net)`
Truncate w	`w_bar[i] + psi * (w_test[i] - w_bar[i])`

Sources

Four GANs side by side

Factor	DCGAN	Pix2Pix	CycleGAN	StyleGAN
Input	noise z	image x, dropout as noise	image from domain X	z mapped to w
Generator	transposed convs, tanh out	U-Net with skips	G: X to Y and F: Y to X	constant plus AdaIN per block
Extra loss	none	λ times L1	cycle consistency	path-length reg in StyleGAN2
Data	images only	paired (x, y)	unpaired	narrow domain, faces
Use for	32x32 baseline	sketch to photo	horses to zebras, seasons	photoreal faces, editing

Sources

One pass against many steps

Factor	Pix2Pix	SD-Img2Img	SDXL-Turbo	ControlNet + SDXL
Steps	1	20	1-4	20-30
512² on one L4	~30 ms	~1.2 s	~0.15-0.35 s	~3-5 s

Sources

Defaults

Setting	Default	Note
Adam `lr`	2e-4	DCGAN default for G and D
Adam `betas`	(0.5, 0.999)	Low beta1
`z_dim`	64	Noise for the 32x32 DCGAN
D LeakyReLU slope	0.2	All D layers
Pix2Pix λ on L1	100	Start here, anneal down
PatchGAN field	~70×70	Pixels per output cell
Truncation ψ	0.7	Use 1.0 for diversity
TTUR ratio	2-4	D learns faster than G

Sources

Troubleshooting

Symptom	Cause	Fix
D loss low, G loss oscillates, variety drops	Mode collapse	Minibatch discrimination, spectral norm, labels
D accuracy ~100%, G loss huge and static	D wins	Smaller D, lower D learning rate, label smoothing
Both losses swing with no trend	Oscillation	TTUR or Wasserstein loss
D above 95% accuracy	D too strong	Cut D learning rate by 2-5x
One real mode never appears	G memorizes a mode	Noise on D inputs, or WGAN-GP
Output ignores the condition	Weak condition signal in D	Condition D early, projection D
Blurry Pix2Pix output	L1 weight too high	Anneal λ down once stable
Droplet artifacts	AdaIN zeroes the mean	Weight demodulation of StyleGAN2

Sources

Terms

DCGAN: strided convs, batch norm, ReLU in G, LeakyReLU in D.
WGAN-GP: Wasserstein distance plus gradient penalty.
SN: spectral norm. Bounds the Lipschitz constant of each D layer.
TTUR: two-timescale update rule. Two learning rates.
cGAN: conditional GAN. G(z, c) and D(x, c) both see the condition.
PatchGAN: D outputs an N × N grid of per-patch scores.
AdaIN: adaptive instance norm. Normalize, then scale and shift by w.
FID: Fréchet Inception Distance. Lower is better.

Sources

Use X when

GAN real-time generation under 10 ms, style transfer, image-to-image translation.
StyleGAN3 photoreal human faces with a fixed pose.
Pix2Pix thousands of paired examples, a narrow task and fast inference.
CycleGAN image translation with no paired data.
Frozen mapping network domain adaptation from a few images.
Diffusion or flow matching anything multi-modal or open-ended.

Sources

Questions

Why G(z).detach() in the D step?: It stops gradients into G during the D update.
Why the non-saturating loss?: Early on D(G(z)) is near zero. The original G loss then has vanishing gradients.
Why switch to eval mode before sampling?: Batch norm then uses running stats.
Why does D see the pair (x, y)?: Without x, D cannot check consistency.
Why does StyleGAN still ship?: One forward pass. Under 10 ms for a 1024² face on a 4090.
Can a GAN give p(x)?: No. It is an implicit density model. It only samples.

Sources

Original source

---
id: gans
title: GANs from DCGAN to StyleGAN
code: GA
subtitle: Losses, DCGAN code, Pix2Pix, CycleGAN and StyleGAN. Defaults, failure signatures and fixes.
hue: 12
sources: phases/08-generative-ai/01-generative-models-taxonomy-history, phases/08-generative-ai/03-gans-generator-discriminator, phases/08-generative-ai/04-conditional-gans-pix2pix, phases/08-generative-ai/05-stylegan, phases/04-computer-vision/09-image-generation-gans
source_hash: bcbf16df3147b7bf
---

## StyleGAN generator
type: architecture
from: phases/08-generative-ai/05-stylegan/docs/en.md

```arch
node | z | 1,1 | z | N(0, I), 512 dims | store
node | map | 2,1 | Mapping network | 8-layer MLP | model
node | w | 3,1 | w in W | roughly disentangled | store
node | const | 1,2 | Learned constant | 4×4×512 | store
node | synth | 2,2 | Synthesis block | upsample, conv, AdaIN, noise | gpu
node | img | 3,2 | Image | 4 up to 1024 | client
node | noise | 1,3 | Per-layer noise | learned per-channel scale | process
edge | z | map |
edge | map | w |
edge | w | synth | AdaIN at every block
edge | const | synth |
edge | noise | synth |
edge | synth | img | resolution doubles
```

## One training step
type: flow
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/03-gans-generator-discriminator/docs/en.md, phases/08-generative-ai/03-gans-generator-discriminator/code/main.py

1. **Sample**: draw a real batch and noise `z` from N(0, I).
2. **D step**: score real images against 1 and `G(z).detach()` against 0.
3. **G step**: make fresh fakes. Score `D(G(z))` against 1.
4. **Repeat**: one step of D, then one step of G, every batch.
5. **Probe**: every 100 steps, count samples per mode. Warn below 50.
6. **Inspect**: look at 64 samples at the end of every epoch.
7. **Equilibrium**: `p_G = p_data` and D outputs 0.5 everywhere.

## Losses and formulas
type: concepts
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/03-gans-generator-discriminator/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md

- **Minimax**: $\min_G \max_D \; \mathbb{E}_x[\log D(x)] + \mathbb{E}_z[\log(1 - D(G(z)))]$
- **D loss**: binary cross-entropy with real = 1 and fake = 0.
- **G loss**: $-\log D(G(z))$. This is the non-saturating form.
- **WGAN**: $\text{loss}_D = \mathbb{E}[D(\text{fake})] - \mathbb{E}[D(\text{real})]$, $\text{loss}_G = -\mathbb{E}[D(\text{fake})]$
- **Pix2Pix G**: $-\log D(x, G(x)) + \lambda \cdot \lVert y - G(x) \rVert_1$
- **Cycle**: $F(G(x)) \approx x$ and $G(F(y)) \approx y$.
- **AdaIN**: $y_{\text{scale}} \cdot \frac{x - \mathrm{mean}(x)}{\mathrm{std}(x)} + y_{\text{bias}}$
- **Truncation**: $w' = \hat{w} + \psi \cdot (w - \hat{w})$. $\psi < 1$ trades diversity for quality.

## DCGAN train_step
type: annotated-code
lang: python
from: phases/04-computer-vision/09-image-generation-gans/code/main.py, phases/04-computer-vision/09-image-generation-gans/docs/en.md

```python
opt_d.zero_grad()
d_real = D(real)                    # logits, sigmoid only in the loss
d_fake = D(G(z).detach())           # no gradient flows into G
loss_d = (F.binary_cross_entropy_with_logits(d_real, torch.ones_like(d_real))
          + F.binary_cross_entropy_with_logits(d_fake, torch.zeros_like(d_fake)))
loss_d.backward()
opt_d.step()
opt_g.zero_grad()
d_fake = D(G(z))                    # second pass keeps the graph
loss_g = F.binary_cross_entropy_with_logits(d_fake, torch.ones_like(d_fake))
loss_g.backward()                   # non-saturating G loss
opt_g.step()
```

## Tasks in code
type: task-table
from: phases/04-computer-vision/09-image-generation-gans/code/main.py, phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/code/main.py, phases/08-generative-ai/05-stylegan/code/main.py

| Task | Code |
|---|---|
| Build the generator | `Generator(z_dim=64, img_channels=3, feat=32)` |
| Wrap a D conv in SN | `spectral_norm(c)` |
| Make the optimizers | `Adam(lr=2e-4, betas=(0.5, 0.999))` |
| Sample in eval mode | `sample(G, n=8, z_dim=z_dim, device=device)` |
| Count parameters | `sum(p.numel() for p in G.parameters())` |
| Condition G or D | `inp = z + one_hot(c, num_classes)` |
| Map z to w | `w = mapping(z, mapping_net)` |
| Truncate w | `w_bar[i] + psi * (w_test[i] - w_bar[i])` |

## Four GANs side by side
type: comparison
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md
span: 2

| Factor | DCGAN | Pix2Pix | CycleGAN | StyleGAN |
|---|---|---|---|---|
| Input | noise z | image x, dropout as noise | image from domain X | z mapped to w |
| Generator | transposed convs, tanh out | U-Net with skips | G: X to Y and F: Y to X | constant plus AdaIN per block |
| Extra loss | none | λ times L1 | cycle consistency | path-length reg in StyleGAN2 |
| Data | images only | paired (x, y) | unpaired | narrow domain, faces |
| Use for | 32x32 baseline | sketch to photo | horses to zebras, seasons | photoreal faces, editing |

## One pass against many steps
type: comparison
from: phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md

| Factor | Pix2Pix | SD-Img2Img | SDXL-Turbo | ControlNet + SDXL |
|---|---|---|---|---|
| Steps | 1 | 20 | 1-4 | 20-30 |
| 512² on one L4 | ~30 ms | ~1.2 s | ~0.15-0.35 s | ~3-5 s |

## Defaults
type: defaults
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/04-computer-vision/09-image-generation-gans/code/main.py, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md

| Setting | Default | Note |
|---|---|---|
| Adam `lr` | 2e-4 | DCGAN default for G and D |
| Adam `betas` | (0.5, 0.999) | Low beta1 |
| `z_dim` | 64 | Noise for the 32x32 DCGAN |
| D LeakyReLU slope | 0.2 | All D layers |
| Pix2Pix λ on L1 | 100 | Start here, anneal down |
| PatchGAN field | ~70×70 | Pixels per output cell |
| Truncation ψ | 0.7 | Use 1.0 for diversity |
| TTUR ratio | 2-4 | D learns faster than G |

## Troubleshooting
type: troubleshooting
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/03-gans-generator-discriminator/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md

| Symptom | Cause | Fix |
|---|---|---|
| D loss low, G loss oscillates, variety drops | Mode collapse | Minibatch discrimination, spectral norm, labels |
| D accuracy ~100%, G loss huge and static | D wins | Smaller D, lower D learning rate, label smoothing |
| Both losses swing with no trend | Oscillation | TTUR or Wasserstein loss |
| D above 95% accuracy | D too strong | Cut D learning rate by 2-5x |
| One real mode never appears | G memorizes a mode | Noise on D inputs, or WGAN-GP |
| Output ignores the condition | Weak condition signal in D | Condition D early, projection D |
| Blurry Pix2Pix output | L1 weight too high | Anneal λ down once stable |
| Droplet artifacts | AdaIN zeroes the mean | Weight demodulation of StyleGAN2 |

## Terms
type: glossary
from: phases/08-generative-ai/03-gans-generator-discriminator/docs/en.md, phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md

- **DCGAN**: strided convs, batch norm, ReLU in G, LeakyReLU in D.
- **WGAN-GP**: Wasserstein distance plus gradient penalty.
- **SN**: spectral norm. Bounds the Lipschitz constant of each D layer.
- **TTUR**: two-timescale update rule. Two learning rates.
- **cGAN**: conditional GAN. G(z, c) and D(x, c) both see the condition.
- **PatchGAN**: D outputs an N × N grid of per-patch scores.
- **AdaIN**: adaptive instance norm. Normalize, then scale and shift by w.
- **FID**: Fréchet Inception Distance. Lower is better.

## Use X when
type: tips
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/03-gans-generator-discriminator/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md, phases/08-generative-ai/01-generative-models-taxonomy-history/docs/en.md

- **GAN**: real-time generation under 10 ms, style transfer, image-to-image translation.
- **StyleGAN3**: photoreal human faces with a fixed pose.
- **Pix2Pix**: thousands of paired examples, a narrow task and fast inference.
- **CycleGAN**: image translation with no paired data.
- **Frozen mapping network**: domain adaptation from a few images.
- **Diffusion or flow matching**: anything multi-modal or open-ended.

## Questions
type: qa
from: phases/04-computer-vision/09-image-generation-gans/docs/en.md, phases/08-generative-ai/04-conditional-gans-pix2pix/docs/en.md, phases/08-generative-ai/05-stylegan/docs/en.md, phases/08-generative-ai/01-generative-models-taxonomy-history/docs/en.md

- **Why `G(z).detach()` in the D step?** It stops gradients into G during the D update.
- **Why the non-saturating loss?** Early on `D(G(z))` is near zero. The original G loss then has vanishing gradients.
- **Why switch to eval mode before sampling?** Batch norm then uses running stats.
- **Why does D see the pair (x, y)?** Without x, D cannot check consistency.
- **Why does StyleGAN still ship?** One forward pass. Under 10 ms for a 1024² face on a 4090.
- **Can a GAN give p(x)?** No. It is an implicit density model. It only samples.