Generative Models
From latent variables, variational inference, and VAEs to diffusion, continuous flows, and flow matching, with image and geometry examples.
Collection contents 11 articles
- From Latent Variable Models to Variational Inference
- Neural Networks as Probability Models
- Trainable Variational Inference: Sampling, Reparameterization, and Amortization
- VAE: From a Latent Probability Model to Generation
- Diffusion Models I: From Corruption and Denoising to Generation
- Diffusion Models II: A Compact Derivation from Corruption to Sampling
- Diffusion Models III: DDPM and DDIM Share Training, Not Sampling Rules
- Continuous-Time Generation I: One Score, SDE or ODE Sampling
- Continuous-Time Generation II: Invertible Maps and Normalizing Flows
- Flow Matching: Learning Motion from Noise to a Chair
- Geometry Generation: ShapeNet Chairs as Point Clouds or Fields
The first four articles cover probability modeling and VAEs. Start diffusion at article five: the first article explains corruption, training, generation, and common extensions using MNIST and ShapeNet chairs. The second gives the complete DDPM mathematical chain with a compact derivation. Continue to faster sampling, continuous time, Flow Matching, and SDF geometry.