Michele LaurelliBlog

DDIM (Denoising Diffusion Implicit Model)

DDIM (Denoising Diffusion Implicit Model): A diffusion sampling method that enables non-Markovian trajectories and fewer steps. It uses a family of implicit processes compatible with the trained model. A parameter controls stochasticity and the trade-off among speed, diversity and quality.

In italiano: DDIM (Denoising Diffusion Implicit Model)Legge II · Probabilità, softmax e modelli generativi

It uses a family of implicit processes compatible with the trained model. A parameter controls stochasticity and the trade-off among speed, diversity and quality.

Examples

  • Sampling in 20 steps
  • Deterministic trajectory
  • Accelerating a DDPM

Fonti