cDDPM in PyTorch

Implementing classifier-free diffusion guidance and exploring U-Net + attention designs for diffusion models.

cDDPM is a PyTorch implementation of classifier-free diffusion guidance, with experiments on how U-Net depth, width, and attention affect the trade-off between sample quality and diversity.

🔍 Summary

Instead of using a separate classifier at sampling time, classifier-free guidance trains a single diffusion backbone in two modes—conditional and unconditional—and combines their scores with a tunable guidance scale. This project:

  • Implements a DDPM-style backbone with classifier-free guidance.
  • Exposes the guidance weight as a simple knob for quality vs. diversity.
  • Uses the implementation as a sandbox to study architectural tweaks.

🧠 Architecture & experiments

  • U-Net backbone with configurable:
    • Number of channels per stage.
    • Attention blocks at chosen resolutions.
  • Experiments on:
    • Varying guidance scales and observing mode collapse vs. sharpness.
    • Changing attention configuration (heads, resolutions) and U-Net depth.

đź“‚ GitHub repository

  • Code & training scripts: github.com/manhbeo/cDDPM
  • Reference: Classifier-Free Diffusion Guidance (Ho & Salimans, 2022, arXiv:2207.12598).