Architectures de modèles
Generative Adversarial Networks (GANs)
Train a generator to fool a discriminator while the discriminator learns to distinguish generated samples from training data.
Modèle mental
A forger and a critic improve against each other; after training, the forger can generate in one forward pass.
Flux de données
- Random latent + condition
- Generator
- Synthetic sample
- Discriminator compares real vs synthetic
- Adversarial gradients update both
Entraînement
A minimax adversarial objective couples two networks. Practical variants change losses, regularization, normalization, and conditioning to stabilize training and control outputs.
Exécution de l’inférence
Sample a latent and run the generator once, making generation fast. Inversion or editing requires additional machinery because there is no built-in iterative reverse process.
Atouts
- Fast one-pass sampling
- Sharp outputs in well-scoped visual domains
- Useful conditional and image-to-image variants
Compromis
- Training instability and mode collapse
- Coverage and likelihood are hard to assess
- Large open-domain text-conditioned generation has shifted toward other families
À utiliser lorsque
- Low-latency sampling matters in a constrained domain
- A proven GAN pipeline already fits the data and controls
- Diversity and coverage are explicitly evaluated
À éviter ou remettre en question lorsque
- Stable training and broad mode coverage are primary requirements
- The team lacks domain-specific evaluation
- A newer diffusion or autoregressive baseline is not being compared
Familles publiées à titre d’exemple
- • Original GAN formulation
- • StyleGAN research family
- • Conditional image-to-image GANs
Souvent combinée avec
Convolutional or Transformer generatorsEncoders for inversion and editing