Ch 11 — Generative AI
GAN math, diffusion forward/reverse process, VAE ELBO, latent diffusion, and classifier-free guidance
Under the Hood
-
Click play or press Space to begin the deep dive...
Zone AGANs & VAEs — Classic Generative ModelsSteps 1–2
1
smart_toy
GAN Loss
Minimax objective
2
compress
VAE & ELBO
Encoder-decoder + KL
arrow_downward From GANs/VAEs to diffusion
3
Zone BDiffusion Models — Forward & Reverse ProcessSteps 3–5
3
blur_on
Forward Process
q(x\u209c|x\u209c\u208b\u2081) noise schedule
4
auto_fix_high
Reverse Process
ε-prediction training
5
speed
Samplers
DDIM, DPM++, Euler
arrow_downward Move to latent space
6
Zone CLatent Diffusion & Text ConditioningSteps 6–7
6
layers
Latent Diffusion
VAE encoder + U-Net/DiT
7
text_fields
CFG
Classifier-free guidance
arrow_downward Architecture evolution
8
Zone DU-Net vs DiT & Flow MatchingSteps 8–9
8
account_tree
U-Net & DiT
CNN vs Transformer denoiser
9
route
Flow Matching
Straight paths, fewer steps
arrow_downward Video & audio internals
10
Zone EVideo & Audio Generation InternalsStep 10
10
movie
Spacetime DiT
Sora, codec audio