Skip to content

Model Zoo Migration¤

The legacy Artifex model zoo has been removed as a live runtime surface.

artifex.generative_models.zoo remains only as a compatibility boundary that raises a migration error. It does not provide a working preset registry, configuration listing API, or a live zoo model-construction API anymore.

Why It Was Removed¤

The old zoo taught a generic preset and model_class story that no longer matched the supported factory contract. The canonical runtime path is now:

  1. materialize a family-specific typed config
  2. pass that typed config to artifex.generative_models.factory.create_model

That keeps model creation aligned with the actual builders the runtime supports.

Supported Replacement¤

Use a concrete config family such as VAEConfig, DDPMConfig, WGANConfig, or PointCloudConfig.

from flax import nnx

from artifex.generative_models.core.configuration import (
    DecoderConfig,
    EncoderConfig,
    VAEConfig,
)
from artifex.generative_models.factory import create_model

encoder = EncoderConfig(
    name="vae_encoder",
    input_shape=(28, 28, 1),
    latent_dim=32,
    hidden_dims=(256, 128),
)
decoder = DecoderConfig(
    name="vae_decoder",
    latent_dim=32,
    output_shape=(28, 28, 1),
    hidden_dims=(128, 256),
)
config = VAEConfig(
    name="mnist_vae",
    encoder=encoder,
    decoder=decoder,
    kl_weight=1.0,
)

model = create_model(config, rngs=nnx.Rngs(0))

Migrating Existing Presets¤

If your project used named zoo presets, move that lookup into your own project code and make it return typed configs.

Recommended pattern:

  • keep preset names in your application layer
  • map each preset name to a typed config constructor or YAML loader
  • call create_model(config, rngs=...) after materializing that typed config

YAML And Template Guidance¤

If you still want reusable presets, keep them as project-local typed config documents and load them with the matching config class. For example:

  • VAEConfig.from_dict(...)
  • DDPMConfig.from_dict(...)
  • PointCloudConfig.from_dict(...)

Do not rebuild a generic universal preset registry around model_class strings.

Summary¤

  • the old zoo runtime has been removed
  • the supported creation path is typed config plus factory

  • project-specific presets should stay outside artifex.generative_models.zoo

  • ❌ Modify zoo configs directly (use overrides instead)

  • ❌ Use generic names that might conflict