Unet-Denoising-Diffusion-Probabilistic-Models

• The code comes from Labml, a very clear DDPM implementation, I moved the code here mainly for learning, and added some comments to the code.
• The original paper Denoising Diffusion Probabilistic Models, i also highly recommend a very good DDPM explanatory article from Google : Luo C. Understanding diffusion models: A unified perspective .

---

• The main structure of the model
  • Unet : Used to generate ${ϵ}_t$ , which is based on $x_t$ and a random index $t$ step that for get the time step embedding .
  • For $x_t$ , you can get it by flowing formula :
    • $q(x_t|x_0)=\mathcal{N}(x_t;\sqrt{{\overline{\alpha }}_t}{x}_0,(1-{\overline{\alpha }}_t)\mathbf{I})$
  • The loss
    • 
  • For more details of the model, please refer to
    • [1] Labml
    • [2] Denoising Diffusion Probabilistic Models
    • [3] Luo C. Understanding diffusion models: A unified perspective