大纲

We will look at

• The corruption process (adding noise to data)
• What a UNet is, and how to implement an extremely minimal one from scratch
• Diffusion model training
• Sampling theory

Then we'll compare our versions with the diffusers DDPM implementation, exploring

• Improvements over our mini UNet
• The DDPM noise schedule
• Differences in training objective
• Timestep conditioning
• Sampling approaches

步骤

Setup and Imports:

The Data

The Corruption Process(adding noise to data)