Diffusion Model Conversions

A unified interface for mapping between different mathematical representations of diffusion processes, essential for training and sampling from diffusion-based generative models.

When to Use

• Building diffusion-based generative models
• Converting neural network predictions (y1, epsilon) to sampling quantities (flow, drift, score)
• Implementing custom samplers for diffusion models
• Understanding the relationships between different diffusion parameterizations

Key Concepts

Diffusion Model Components

A diffusion process is defined by: