| Diffusion Models |
The forward and reverse processes, noise schedules, the ELBO, training and sampling algorithms, network architectures, and latent diffusion |
Open |
| Stable Diffusion |
The attention U-Net, text conditioning with CLIP, latent diffusion, classifier-free guidance and samplers, and Textual Inversion, DreamBooth, LoRA and ControlNet |
Open |
| Lab 1: DDIM on Anime Faces |
Train a diffusers U-Net on anime faces with the DDPM noise schedule, then sample with the faster DDIM scheduler |
Notebook Colab |
| Lab 2: Latent Diffusion on Anime Faces |
Diffuse in the latent space of a frozen VQ autoencoder with diffusers: the U-Net, noise schedulers, the training objective and sampling |
Notebook Colab |
| Optional: Advanced Image Generation Models |
Cascaded, score-based and PDE-based diffusion, text-conditioned diffusion, and CLIP-guided generation and editing |
Open |
| Optional: Stable Diffusion with diffusers |
Generate images with StableDiffusionPipeline, then write your own inference loop with classifier-free guidance |
Notebook Colab |
| Optional: Score-Based Generative Modeling |
Score matching, diffusion as a stochastic differential equation, and sampling with predictor-corrector and ODE solvers |
Notebook Colab |
| Optional: Flow Matching on Anime Faces |
Flow matching as regressing a velocity field: how it relates to diffusion, a 2-D demonstration, then anime faces in a handful of sampling steps |
Notebook Colab |
