From DDPM to Latent Diffusion and DiT

Forward process

Repeated Gaussian perturbations move clean (x_0) toward a simple (x_T) distribution.

Reverse process

Generation starts from noise and moves from (T) toward zero. The model predicts a distribution parameter or the added noise at each step.

Noise prediction

Training samples a timestep and constructs (x_t) directly with cumulative coefficients. Sampling remains sequential over all timesteps.

Latent diffusion and DiT

Latent diffusion denoises inside an autoencoder representation and decodes the image afterward. DiT replaces the U-Net noise predictor with a transformer over latent patches.

Rerun order

The course notebook retained code but no outputs. I will first check tensor shapes and the sampling loop on FashionMNIST, then add the latent encoder and decoder boundary.