Diffusion

Module: artifex.generative_models.core.sampling.diffusion

Source: src/artifex/generative_models/core/sampling/diffusion.py

Overview

DiffusionSampler retains two supported responsibilities:

• stateful DDPM-style stepping through init(...) and step(...)
• a wrapper-only sample(...) entrypoint that delegates to model.sample(...)

The public sample(...) method does not implement a standalone generic direct-sampling path. If you want DiffusionSampler.sample(...), initialize the sampler with a model that already owns a real sample(...) implementation.

Supported Sampling Contract

• DiffusionSampler.sample(...) is wrapper-only.
• It delegates to model.sample(...) when the sampler was created with a compatible model owner.
• It forwards scheduler, optional steps, and optional rngs to that model.
• Without a model-owned sampling implementation, the method raises NotImplementedError instead of pretending a generic fallback exists.

Stepper Contract

init(...) and step(...) still provide the retained low-level diffusion stepping utility for explicit state dictionaries containing x, key, and t. Use this surface when you want to own the outer loop yourself.