---

name: gan description: "Routes legacy GAN-family MNIST scripts, framework choices, model variant lookup, and TensorFlow/PyTorch compatibility troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Unlicense

GAN

Use this sub-skill for GAN-family requests in the Generative Models repository: vanilla GAN, conditional GAN, InfoGAN, WGAN, WGAN-GP, LSGAN, ACGAN, BEGAN, BGAN, EBGAN, f-GAN, GAP, DiscoGAN, DualGAN, COGAN, ALI/BiGAN, MAGAN, Softmax GAN, GibbsNet, and mode-regularized GAN.

Route here when

• The user asks which GAN variant or source artifact label matches a modeling goal.
• The user needs TensorFlow-vs-PyTorch guidance for a legacy GAN example.
• The user asks about GAN sample outputs, MNIST assumptions, or why a GAN script fails on a modern stack.
• The user names WGAN-GP, InfoGAN, DiscoGAN, DualGAN, ALI/BiGAN, GibbsNet, or another GAN variant from the catalog.

Route elsewhere when

• VAE / CVAE / AAE / AVB: ../vae/SKILL.md
• Binary RBM CD/PCD: ../rbm/SKILL.md
• Helmholtz Machine wake-sleep: ../helmholtz-machine/SKILL.md
• Shared catalog, provenance, and compatibility: ../../SKILL.md

Fast decisions

• Default GAN baseline: vanilla GAN.
• Class-conditioned generation: conditional GAN or ACGAN.
• Wasserstein objective with gradient penalty: WGAN-GP, TensorFlow-only in this repo.
• Image-to-image / domain translation style examples: DiscoGAN, DualGAN, or COGAN.
• Bidirectional inference: ALI/BiGAN or GibbsNet.
• Mode-collapse or divergence variants: mode-regularized GAN, MAGAN, f-GAN, LSGAN, Softmax GAN, BGAN, BEGAN, or EBGAN.

Read next

• references/workflows.md for the full GAN variant map and framework coverage.
• references/troubleshooting.md for GAN-specific modern-stack failures.
• ../../references/model-catalog.md for the repo-wide source artifact catalog.
• ../../references/compatibility.md for the shared TensorFlow/PyTorch/NumPy compatibility matrix.
• ../../scripts/check_legacy_stack.py before claiming an unchanged GAN example can execute.