Skip to content
LearnMLByDoing
Free
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

  1. 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
  2. 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
  3. 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
  4. 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