DDIM

Denoising Diffusion Implicit Models

Jiaming Song, Chenlin Meng, Stefano Ermon · ICLR 2021

Read paper · arXiv:2010.02502

Use a non-Markovian sampling update to skip timesteps, with eta controlling stochasticity.

What you will implement

Implement the DDIM update and clean endpoint; timestep selection, loops, and inversion exploration code are supplied.

• 9. DDIM Subsequence Update

  Implement one DDIM update, combining the clean prediction, direction, and stochastic variance correctly.

  Implement ddim_step only. Timestep selection and sampling are provided; inversion is an optional exploration helper.