Generative models beyond text
Learn reconstruction, latent representations and the limits of sampling from small trained models.
What you’ll explore
- Autoencoders
- Reconstruction
- Latent sampling
- Variational inference
- ELBO
- Mixture models
- Expectation–maximization
- Diffusion
- Noise prediction
Your lessons
0 / 8 completeCompress and reconstruct
Train a tiny shape autoencoder, inspect its bottleneck and test why reconstruction does not guarantee useful generation.
Generate with a variational autoencoder
Train a tiny probabilistic encoder and decoder, trace a noisy latent sample and compare reconstruction with KL regularization.
Infer what cannot be directly observed
Fit a two-component mixture in separate EM steps and compare exact latent inference with a restricted variational approximation.
Generate by removing noise
Train a small 2D noise predictor, inspect corruption arithmetic and manually step a frozen DDPM sampler.
Generate by learning a flow
Train a velocity field and compare Euler and Heun generation on the diffusion lesson’s four-mode target.
Control the generated result
Train a shape-conditioned flow and compare guidance, validity and coverage at equal sampling work.
Learn with a generator and discriminator
Train an actual tiny GAN, inspect discriminator gradients and compare missing modes with a reproducible diffusion reference.
Judge a generated distribution
Compare fidelity, mode coverage, copying flags, task gates and licensed source receipts instead of trusting one score.