DDIM

Overview

Generalize DDPMs from Markovian to non-Markovian forward processes. The training objective is actually the same. This improves:

• Generated Quality
• Consistency property - if we generate using a different number of steps, we get similar high-level features.
• Semantically meaningful image interpolation via latent variable interpolation.

Derivation

Note that the DDPM objective depends only on the marginal distributions \(q(\mathbf{x_t} \mid \mathbf{x_0})\) and not \(q(\mathbf{x_t} \mid \mathbf{x_0}, ... , \mathbf{x_T})\)