---

name: generative-models description: "Routes Papers-in-100-Lines GAN, VAE, flow, diffusion, DreamBooth, Stable Diffusion, and image-translation implementation tasks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Generative Models

Use this sub-skill when the user asks about generative implementations in Papers-in-100-Lines: GAN variants, VAEs, normalizing flows, diffusion samplers, DreamBooth, Stable Diffusion, image translation, SNL/AALR, or related toy density/image generation examples.

Read these bundled files

• Model family guide maps the bundled catalog entries to model families, key classes/functions, data assumptions, and safe adaptation recipes.
• Troubleshooting covers dataset downloads, old dependency pins, CUDA hardcodes, missing Stable Diffusion weights, output-image side effects, and shape/loss failures.
• Implementation index lists all entries; use it when the user gives only a paper title or acronym.
• Dependency and backend guide explains why each paper should normally get its own environment.
• summarize_generative_entries.py summarizes this family's catalog entries without importing torch, Keras, or Diffusers.

Trigger routes

• GAN or image translation: original GAN, CGAN, DCGAN, LSGAN, WGAN, WGAN-GP, Improved GAN, ALI/AFL, pix2pix, CycleGAN.
• VAE, flow, or density estimation: Auto-Encoding Variational Bayes, NICE, Real NVP, MAF, planar flows, SNL, AALR-MCMC, Gromov-Wasserstein.
• Diffusion and text-to-image: nonequilibrium thermodynamics, DDPM, DDIM, PNDM, DPM-Solver, rectified flow, Stable Diffusion v1-5, DreamBooth.
• Safety triage: detect when the requested action would download MNIST, tokenizers, model weights, or run a long GPU training loop.

Workflow for a future agent

1. If the paper is not explicit, query the catalog first:

   python ../../scripts/query_implementation_index.py --group generative-models --query "diffusion"
   

2. Read Model family guide for the selected family and identify the reusable pattern: generator/discriminator, encoder and latent distribution, invertible transform, score/noise predictor, or sampler schedule.

3. Decide whether the user needs full reproduction or a small adaptation. For small adaptation, create a new tiny script with explicit device, data, and output arguments rather than importing a file that may download data at import time.

4. Use per-entry requirements from the catalog. Do not merge all generative requirements into one environment.

5. Validate with a bounded shape or loss check before training: generator output shape, flow log-prob shape, diffusion sample tensor shape, or text-to-image weight/tokenizer presence.

Boundaries

Route NeRF, implicit representation, splatting, camera/ray, and 3D reconstruction tasks to neural-rendering-3d. Route optimizers, activations, meta-learning, hypergradients, Deep Image Prior, and Atari RL tasks to optimization-meta-rl. Use paper-catalog-and-execution for lookup and first-run safety planning.