Diffusion Model

 Diffusion Model can train probability distribution of given dataset effectively.

It can understand geometry of target well.

https://youtu.be/XZ0PMRWXBEU?si=GzOrWxuX25dZieWR

Generative Model: Variational Inference, Variational Auto-Encoder (VAE)

Markovian Hierarchical VAE: Denoising Diffusion Probabilistic Models (DDPM)

Connect with Langevin Dynamics: Score-Based Model

non-Markovian Hierarchical VAE: Denoising Diffusion Implicit Models (DDIM)

Guidance: CFG & ControlNet

3D-Object Generation: Score Distillation Sampling (SDS)

(With pre-trained noise predictor, use many kind of denoising sampler.)

DPM-solver-0 == linear == DDPM

DPM-solver-1 == DDIM

DPM-solver-n == n-th order approximation (Feynmann Diagram)