May 2026 Course -Lecture 4: DDPM, the Modern Starting Point of Diffusion Image Generation

2038 pages, including 194 video tutorials.

Core paper:

Denoising Diffusion Probabilistic Models
Jonathan Ho, Ajay Jain, Pieter Abbeel, 2020

Why it is essential:

DDPM turned generation into a multi-step denoising process and showed high-quality image synthesis. It also connected diffusion probabilistic models with denoising score matching and Langevin dynamics.

Topics:

• Forward noising process

• Reverse denoising process

• Timestep embeddings

• Noise schedule

• Noise prediction objective

• Sampling cost vs image quality

Key concept:

Training:
x0 + noise → xt
model predicts noise

Generation:
random noise xT → xT-1 → ... → x0

Engineering connection:

• Stable Diffusion
  

• Video diffusion
  

• Seeds
  

• Sampling steps
  

• Schedulers