How diffusion models generate images: noise and denoise, DDPM and fast samplers, U-Net vs DiT, latent diffusion, classifier-free guidance and flow matching.