05
Generative AI and Diffusion
Learn to reverse noise: score fields, Langevin dynamics and the denoising networks behind modern image generators.
- Lessons
- 4
- Study time
- 2.5 h
- Access
- Pro track
Syllabus
- 5.1
Autoencoders, VAEs and latent spaces
Compress, sample, decode: the ELBO, the reparameterisation trick and smooth latent manifolds.
36 minProLive 3D lab and Python challengeadvanced- Derive the evidence lower bound
- Apply the reparameterisation trick
- Interpolate through a learned latent space
- 5.2
Score matching and Langevin dynamics
The gradient of log-density as a vector field, and sampling by following it with injected noise.
38 minProLive 3D lab and Python challengeadvanced- Define the score function ∇ₓ log p(x)
- Train a score model with denoising score matching
- Sample with annealed Langevin dynamics
- 5.3
DDPM: the forward and reverse processes
A fixed noising Markov chain, a learned denoiser, and the simple ε-prediction loss that ties them together.
42 minProLive 3D lab and Python challengeadvanced- Derive the closed-form forward marginal q(xₜ | x₀)
- Train an ε-prediction network
- Implement the ancestral sampling loop
- 5.4
Classifier-free guidance and latent diffusion
Steering samples with conditioning, and running diffusion in a compressed latent space for speed.
34 minProLive 3D lab and Python challengeadvanced- Combine conditional and unconditional scores
- Trade diversity for fidelity with the guidance scale
- Explain the latent diffusion pipeline