Sampling Methods¤

Sampling in Artifex stays family-owned. The runtime does not ship a second generic sampler framework on top of the model families, so the supported entry points are the methods owned by each retained model package.

Live Sampling Surface¤

Family	Live owners	Retained sampling entry points
VAE	`artifex.generative_models.models.vae`	`sample(...)`, `reconstruct(...)`, `decode(...)`
GAN	`artifex.generative_models.models.gan`	`generate(...)`
Diffusion	`artifex.generative_models.models.diffusion`	`generate(...)`, `DDPMModel.sample(..., scheduler="ddim")`
Flow	`artifex.generative_models.models.flow`	`sample(...)`, `generate(...)`, `log_prob(...)`

There is no exported FlowModel alias and no standalone DDIMSampler, DPMSolver, or StyleGANSampler helper in the current runtime.

Imports¤

import jax
from flax import nnx

from artifex.generative_models.models.diffusion import DDPMModel
from artifex.generative_models.models.flow import NormalizingFlow
from artifex.generative_models.models.gan import GAN
from artifex.generative_models.models.vae import VAE

VAE Sampling And Reconstruction¤

def sample_vae(vae: VAE, batch: jax.Array) -> tuple[jax.Array, jax.Array]:
    samples = vae.sample(8, temperature=0.8)
    reconstructions = vae.reconstruct(batch, deterministic=True)
    return samples, reconstructions

GAN Generation¤

def sample_gan(gan: GAN) -> jax.Array:
    return gan.generate(8)

Diffusion Generation¤

Use the retained family owner directly. generate(...) keeps the default DDPM path, while DDPMModel.sample(..., scheduler="ddim") is the live fast-path entry point when you want fewer denoising steps.

def sample_ddpm(ddpm: DDPMModel) -> tuple[jax.Array, jax.Array]:
    full_schedule = ddpm.generate(4)
    fast_schedule = ddpm.sample(4, scheduler="ddim", steps=50)
    return full_schedule, fast_schedule

Flow Sampling And Scoring¤

Flows keep both generation and density evaluation on the model itself.

def sample_flow(flow: NormalizingFlow, batch: jax.Array) -> tuple[jax.Array, jax.Array]:
    samples = flow.sample(8, rngs=nnx.Rngs(3))
    log_probs = flow.log_prob(batch, rngs=nnx.Rngs(4))
    return samples, log_probs

Practical Rule¤

If a sampling example needs a new wrapper class, exported helper namespace, or cross-family orchestration object, check the runtime owner first. Keep the docs on the family method that already exists unless a real shared inference owner is added to the codebase.